OpenMP Application Program Interface Version 3.1 July 2011 Copyright © 1997-2011 OpenMP Architecture Review Board. Permission to copy without fee all or part of this material is granted, provided the OpenMP Architecture Review Board copyright notice and the title of this document appear. Notice is given that copying is by permission of OpenMP Architecture Review Board. This page intentionally left blank. C O N T E N TS 1. 2. Introduction ...............................................1 1.1 Scope ................................................1 1.2 Glossary ..............................................2 1.2.1 Threading Concepts 1.2.2 OpenMP Language Terminology 1.2.3 Tasking Terminology 1.2.4 Data Terminology 1.2.5 Implementation Terminology 1.3 Execution Model 1.4 Memory Model ..............................2 .....................2 ..............................8 .................................9 . . . . . . . . . . . . . . . . . . . . . . . . 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.4.1 Structure of the OpenMP Memory Model 1.4.2 The Flush Operation 1.4.3 OpenMP Memory Consistency . . . . . . . . . . . . . . . 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 . . . . . . . . . . . . . . . . . . . . . . 16 1.5 OpenMP Compliance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 1.6 Normative References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 1.7 Organization of this document Directives 2.1 2.2 2.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Directive Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.1.1 Fixed Source Form Directives . . . . . . . . . . . . . . . . . . . . . . . 23 2.1.2 Free Source Form Directives . . . . . . . . . . . . . . . . . . . . . . . . 24 Conditional Compilation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.2.1 Fixed Source Form Conditional Compilation Sentinels 2.2.2 Free Source Form Conditional Compilation Sentinel . . . . 26 . . . . . . 27 Internal Control Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 ICV Descriptions i 2.4 2.3.2 Modifying and Retrieving ICV Values 2.3.3 How the Per-Data Environment ICVs Work 2.3.4 ICV Override Relationships parallel Construct 2.4.1 2.5 2.6 2.7 2.8 2.9 ii . . . . . . . . . . . . . . . . . . 29 . . . . . . . . . . . . . 30 . . . . . . . . . . . . . . . . . . . . . . . . . 31 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 Determining the Number of Threads for a parallel Region 36 Worksharing Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2.5.1 Loop Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 2.5.2 sections Construct 2.5.3 single Construct 2.5.4 workshare Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 Combined Parallel Worksharing Constructs 2.6.1 Parallel Loop Construct 2.6.2 parallel sections Construct 2.6.3 parallel workshare Construct Tasking Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 . . . . . . . . . . . . . . . . . . . . . 57 . . . . . . . . . . . . . . . . . . . . 59 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 2.7.1 task Construct 2.7.2 taskyield Construct 2.7.3 Task Scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 Master and Synchronization Constructs 2.8.1 master Construct 2.8.2 critical Construct 2.8.3 barrier Construct 2.8.4 taskwait Construct 2.8.5 atomic Construct 2.8.6 flush Construct 2.8.7 ordered Construct Data Environment . . . . . . . . . . . . . . . . . . . 55 . . . . . . . . . . . . . . . . . . . . . . 67 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 2.9.1 Data-sharing Attribute Rules 2.9.2 threadprivate Directive 2.9.3 Data-Sharing Attribute Clauses OpenMP API • Version 3.1 July 2011 . . . . . . . . . . . . . . . . . . . . . . . . 84 . . . . . . . . . . . . . . . . . . . . . . . . . 88 . . . . . . . . . . . . . . . . . . . . . . 92 2.9.4 2.10 3. Data Copying Clauses Nesting of Regions Runtime Library Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 3.1 Runtime Library Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 3.2 Execution Environment Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 3.2.1 omp_set_num_threads . . . . . . . . . . . . . . . . . . . . . . . . . . 116 3.2.2 omp_get_num_threads . . . . . . . . . . . . . . . . . . . . . . . . . . 117 3.2.3 omp_get_max_threads . . . . . . . . . . . . . . . . . . . . . . . . . . 118 3.2.4 omp_get_thread_num 3.2.5 omp_get_num_procs 3.2.6 omp_in_parallel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 3.2.7 omp_set_dynamic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 3.2.8 omp_get_dynamic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 3.2.9 omp_set_nested . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 3.2.10 omp_get_nested . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 3.2.11 omp_set_schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 3.2.12 omp_get_schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 3.2.13 omp_get_thread_limit . . . . . . . . . . . . . . . . . . . . . . . . . 131 3.2.14 omp_set_max_active_levels . . . . . . . . . . . . . . . . . . . . 132 3.2.15 omp_get_max_active_levels . . . . . . . . . . . . . . . . . . . . 134 3.2.16 omp_get_level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 3.2.17 omp_get_ancestor_thread_num 3.2.18 omp_get_team_size . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 3.2.19 omp_get_active_level 3.2.20 omp_in_final 3.3 Lock Routines . . . . . . . . . . . . . . . . . . 136 . . . . . . . . . . . . . . . . . . . . . . . . . 139 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 3.3.1 omp_init_lock and omp_init_nest_lock . . . . . . . . . 143 3.3.2 omp_destroy_lock and omp_destroy_nest_lock 3.3.3 omp_set_lock and omp_set_nest_lock . . . 144 . . . . . . . . . . . . 145 iii 3.4 4. 3.3.4 omp_unset_lock and omp_unset_nest_lock 3.3.5 omp_test_lock and omp_test_nest_lock Timing Routines . . . . . . . . . . 147 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 3.4.1 omp_get_wtime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 3.4.2 omp_get_wtick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 Environment Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 4.1 OMP_SCHEDULE 4.2 OMP_NUM_THREADS 4.3 OMP_DYNAMIC 4.4 OMP_PROC_BIND 4.5 OMP_NESTED 4.6 OMP_STACKSIZE 4.7 OMP_WAIT_POLICY 4.8 OMP_MAX_ACTIVE_LEVELS 4.9 OMP_THREAD_LIMIT A. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 A.1 A Simple Parallel Loop A.2 The OpenMP Memory Model A.3 Conditional Compilation A.4 Internal Control Variables (ICVs) A.5 The parallel Construct A.6 Controlling the Number of Threads on Multiple Nesting Levels A.7 Interaction Between the num_threads Clause and omp_set_dynamic 177 A.8 Fortran Restrictions on the do Construct . . . . . . . . . . . . . . . . . . . . . 179 A.9 Fortran Private Loop Iteration Variables . . . . . . . . . . . . . . . . . . . . . . 181 A.10 The nowait clause A.11 The collapse clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172 OpenMP API • Version 3.1 July 2011 . . . . 175 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 A.12 The parallel sections Construct iv . . . . . . . . 146 . . . . . . . . . . . . . . . . . . . . . . . . 189 A.13 The firstprivate Clause and the sections Construct A.14 The single Construct A.15 Tasking Constructs . . . . . . 190 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 A.16 The taskyield Directive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 A.17 The workshare Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213 A.18 The master Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 A.19 The critical Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219 A.20 worksharing Constructs Inside a critical Construct A.21 Binding of barrier Regions A.22 The atomic Construct . . . . . . . . . . 221 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224 A.23 Restrictions on the atomic Construct A.24 The flush Construct without a List . . . . . . . . . . . . . . . . . . . . . . . 230 . . . . . . . . . . . . . . . . . . . . . . . . . 233 A.25 Placement of flush, barrier, taskwait and taskyield Directives 236 A.26 The ordered Clause and the ordered Construct A.27 The threadprivate Directive . . . . . . . . . . . . . 239 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 A.28 Parallel Random Access Iterator Loop . . . . . . . . . . . . . . . . . . . . . . . 250 A.29 Fortran Restrictions on shared and private Clauses with Common Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 A.30 The default(none) Clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253 A.31 Race Conditions Caused by Implied Copies of Shared Variables in Fortran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 A.32 The private Clause . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256 A.33 Fortran Restrictions on Storage Association with the private Clause 260 A.34 C/C++ Arrays in a firstprivate Clause A.35 The lastprivate Clause A.36 The reduction Clause A.37 The copyin Clause . . . . . . . . . . . . . . . . . . . 263 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 A.38 The copyprivate Clause A.39 Nested Loop Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278 v A.40 Restrictions on Nesting of Regions . . . . . . . . . . . . . . . . . . . . . . . . . . 281 A.41 The omp_set_dynamic and omp_set_num_threads Routines A.42 The omp_get_num_threads Routine A.43 The omp_init_lock Routine A.44 Ownership of Locks . . . . . . . . . . . . . . . . . . . . . . 289 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293 A.45 Simple Lock Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294 A.46 Nestable Lock Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297 B. Stubs for Runtime Library Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301 B.1 C/C++ Stub Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302 B.2 Fortran Stub Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 C. OpenMP C and C++ Grammar C.1 Notation C.2 Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 D. Interface Declarations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325 D.1 Example of the omp.h Header File . . . . . . . . . . . . . . . . . . . . . . . . . 326 D.2 Example of an Interface Declaration include File . . . . . . . . . . . . . 328 D.3 Example of a Fortran Interface Declaration module . . . . . . . . . . . . 330 D.4 Example of a Generic Interface for a Library Routine . . . . . . . . . . . . 334 E. OpenMP Implementation-Defined Behaviors F. Features History . . . . . . . . . . . . . . . . . . . . . 335 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339 F.1 Version 3.0 to 3.1 Differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339 F.2 Version 2.5 to 3.0 Differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340 Index vi . . 288 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343 OpenMP API • Version 3.1 July 2011 1 CHAPTER 1 2 Introduction 3 4 5 6 The collection of compiler directives, library routines, and environment variables described in this document collectively define the specification of the OpenMP Application Program Interface (OpenMP API) for shared-memory parallelism in C, C++ and Fortran programs. 7 8 9 10 This specification provides a model for parallel programming that is portable across shared memory architectures from different vendors. Compilers from numerous vendors support the OpenMP API. More information about the OpenMP API can be found at the following web site 11 http://www.openmp.org 12 13 14 15 16 17 18 19 20 The directives, library routines, and environment variables defined in this document allow users to create and manage parallel programs while permitting portability. The directives extend the C, C++ and Fortran base languages with single program multiple data (SPMD) constructs, tasking constructs, worksharing constructs, and synchronization constructs, and they provide support for sharing and privatizing data. The functionality to control the runtime environment is provided by library routines and environment variables. Compilers that support the OpenMP API often include a command line option to the compiler that activates and allows interpretation of all OpenMP directives. 21 22 23 24 25 26 27 1.1 Scope The OpenMP API covers only user-directed parallelization, wherein the programmer explicitly specifies the actions to be taken by the compiler and runtime system in order to execute the program in parallel. OpenMP-compliant implementations are not required to check for data dependencies, data conflicts, race conditions, or deadlocks, any of which may occur in conforming programs. In addition, compliant implementations are not required to check for code sequences that cause a program to be classified as non1 conforming. Application developers are responsible for correctly using the OpenMP API to produce a conforming program. The OpenMP API does not cover compiler-generated automatic parallelization and directives to the compiler to assist such parallelization. 1 2 3 4 1.2 Glossary 5 1.2.1 Threading Concepts 6 7 8 thread 9 OpenMP thread 10 11 thread-safe routine 12 1.2.2 13 14 15 An execution entity with a stack and associated static memory, called threadprivate memory. A thread that is managed by the OpenMP runtime system. A routine that performs the intended function even when executed concurrently (by more than one thread). OpenMP Language Terminology base language A programming language that serves as the foundation of the OpenMP specification. COMMENT: See Section 1.6 on page 17 for a listing of current base languages for the OpenMP API. 16 17 18 base program 19 20 structured block A program written in a base language. For C/C++, an executable statement, possibly compound, with a single entry at the top and a single exit at the bottom, or an OpenMP construct. 21 22 For Fortran, a block of executable statements with a single entry at the top and a single exit at the bottom, or an OpenMP construct. 23 COMMENTS: 24 For all base languages, 25 • Access to the structured block must not be the result of a branch. 26 27 • The point of exit cannot be a branch out of the structured block. 2 OpenMP API • Version 3.1 July 2011 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 6 7 8 9 • An expression statement, iteration statement, selection statement, or try block is considered to be a structured block if the corresponding compound statement obtained by enclosing it in { and } would be a structured block. 10 For Fortran: 11 12 13 • enclosing context In C/C++, the innermost scope enclosing an OpenMP construct. In Fortran, the innermost scoping unit enclosing an OpenMP construct. 14 15 16 STOP statements are allowed in a structured block. directive In C/C++, a #pragma, and in Fortran, a comment, that specifies OpenMP program behavior. COMMENT: See Section 2.1 on page 22 for a description of OpenMP directive syntax. 17 18 19 white space 20 21 OpenMP program 22 23 conforming program An OpenMP program that follows all the rules and restrictions of the OpenMP specification. 24 25 26 declarative directive An OpenMP directive that may only be placed in a declarative context. A declarative directive has no associated executable user code, but instead has one or more associated user declarations. executable directive An OpenMP directive that is not declarative. That is, it may be placed in an executable context. COMMENT: All directives except the threadprivate directive are executable directives. 30 31 32 A program that consists of a base program, annotated with OpenMP directives and runtime library routines. COMMENT: Only the threadprivate directive is a declarative directive. 27 28 29 A non-empty sequence of space and/or horizontal tab characters. stand-alone directive An OpenMP executable directive that has no associated executable user code. Chapter 1 Introduction 3 loop directive 1 2 An OpenMP executable directive whose associated user code must be a loop nest that is a structured block. COMMENTS: 3 4 For C/C++, only the for directive is a loop directive. 5 6 For Fortran, only the do directive and the optional end do directive are loop directives. associated loop(s) 7 The loop(s) controlled by a loop directive. COMMENT: If the loop directive contains a collapse clause then there may be more than one associated loop. 8 9 10 11 12 13 construct An OpenMP executable directive (and for Fortran, the paired end directive, if any) and the associated statement, loop or structured block, if any, not including the code in any called routines. That is, in the lexical extent of an executable directive. 14 15 16 17 18 19 region All code encountered during a specific instance of the execution of a given construct or of an OpenMP library routine. A region includes any code in called routines as well as any implicit code introduced by the OpenMP implementation. The generation of a task at the point where a task directive is encountered is a part of the region of the encountering thread, but the explicit task region associated with the task directive is not. COMMENTS: 20 21 22 A region may also be thought of as the dynamic or runtime extent of a construct or of an OpenMP library routine. 23 24 During the execution of an OpenMP program, a construct may give rise to many regions. 25 active parallel region 26 27 inactive parallel region 28 4 A parallel region that is executed by a team consisting of more than one thread. A parallel region that is executed by a team of only one thread. OpenMP API • Version 3.1 July 2011 1 2 3 sequential part All code encountered during the execution of an OpenMP program that is not part of a parallel region corresponding to a parallel construct or a task region corresponding to a task construct. COMMENTS: 4 5 6 The sequential part executes as if it were enclosed by an inactive parallel region. 7 8 9 Executable statements in called routines may be in both the sequential part and any number of explicit parallel regions at different points in the program execution. 10 11 master thread The thread that encounters a parallel construct, creates a team, generates a set of tasks, then executes one of those tasks as thread number 0. 12 13 14 15 16 parent thread The thread that encountered the parallel construct and generated a parallel region is the parent thread of each of the threads in the team of that parallel region. The master thread of a parallel region is the same thread as its parent thread with respect to any resources associated with an OpenMP thread. 17 18 ancestor thread 19 20 team For a given thread, its parent thread or one of its parent thread’s ancestor threads. A set of one or more threads participating in the execution of a parallel region. COMMENTS: 21 22 23 For an active parallel region, the team comprises the master thread and at least one additional thread. 24 25 For an inactive parallel region, the team comprises only the master thread. 26 27 28 initial thread implicit parallel region 29 nested construct 30 31 nested region 32 33 The thread that executes the sequential part. The inactive parallel region that encloses the sequential part of an OpenMP program. A construct (lexically) enclosed by another construct. A region (dynamically) enclosed by another region. That is, a region encountered during the execution of another region. COMMENT: Some nestings are conforming and some are not. See Section 2.10 on page 111 for the restrictions on nesting. Chapter 1 Introduction 5 1 2 closely nested region A region nested inside another region with no parallel region nested between them. 3 all threads 4 current team 5 encountering thread 6 all tasks 7 8 9 10 current team tasks All tasks encountered by the corresponding team. Note that the implicit tasks constituting the parallel region and any descendant tasks encountered during the execution of these implicit tasks are included in this binding task set. 11 generating task For a given region the task whose execution by a thread generated the region. 12 13 binding thread set All OpenMP threads participating in the OpenMP program. All threads in the team executing the innermost enclosing parallel region For a given region, the thread that encounters the corresponding construct. All tasks participating in the OpenMP program. The set of threads that are affected by, or provide the context for, the execution of a region. 14 15 The binding thread set for a given region can be all threads, the current team, or the encountering thread. 16 17 COMMENT: The binding thread set for a particular region is described in its corresponding subsection of this specification. binding task set 18 19 The set of tasks that are affected by, or provide the context for, the execution of a region. 20 21 The binding task set for a given region can be all tasks, the current team tasks, or the generating task. 22 23 COMMENT: The binding task set for a particular region (if applicable) is described in its corresponding subsection of this specification. 6 OpenMP API • Version 3.1 July 2011 1 2 binding region The enclosing region that determines the execution context and limits the scope of the effects of the bound region is called the binding region. 3 4 5 Binding region is not defined for regions whose binding thread set is all threads or the encountering thread, nor is it defined for regions whose binding task set is all tasks. 6 COMMENTS: 7 8 The binding region for an ordered region is the innermost enclosing loop region. 9 10 The binding region for a taskwait region is the innermost enclosing task region. 11 12 13 For all other regions for which the binding thread set is the current team or the binding task set is the current team tasks, the binding region is the innermost enclosing parallel region. 14 15 For regions for which the binding task set is the generating task, the binding region is the region of the generating task. 16 17 A parallel region need not be active nor explicit to be a binding region. 18 A task region need not be explicit to be a binding region. 19 20 A region never binds to any region outside of the innermost enclosing parallel region. 21 22 23 24 25 orphaned construct worksharing construct A construct that gives rise to a region whose binding thread set is the current team, but is not nested within another construct giving rise to the binding region. A construct that defines units of work, each of which is executed exactly once by one of the threads in the team executing the construct. 26 For C/C++, worksharing constructs are for, sections, and single. 27 28 For Fortran, worksharing constructs are do, sections, single and workshare. 29 sequential loop 30 31 32 33 barrier A loop that is not associated with any OpenMP loop directive. A point in the execution of a program encountered by a team of threads, beyond which no thread in the team may execute until all threads in the team have reached the barrier and all explicit tasks generated by the team have executed to completion. Chapter 1 Introduction 7 1 1.2.3 Tasking Terminology 2 3 task A specific instance of executable code and its data environment, generated when a thread encounters a task construct or a parallel construct. 4 task region A region consisting of all code encountered during the execution of a task. COMMENT: A parallel region consists of one or more implicit task regions. 5 6 7 explicit task A task generated when a task construct is encountered during execution. 8 9 implicit task A task generated by the implicit parallel region or generated when a parallel construct is encountered during execution. 10 initial task 11 12 current task 13 14 child task 15 16 descendant task A task that is the child task of a task region or of one of its descendant task regions. 17 18 task completion Task completion occurs when the end of the structured block associated with the construct that generated the task is reached. The implicit task associated with the implicit parallel region. For a given thread, the task corresponding to the task region in which it is executing. A task is a child task of the region of its generating task. A child task region is not part of its generating task region. COMMENT: Completion of the initial task occurs at program exit. 19 task scheduling point 20 21 22 A point during the execution of the current task region at which it can be suspended to be resumed later; or the point of task completion, after which the executing thread may switch to a different task region. COMMENT: 23 24 25 Within tied task regions, task scheduling points only appear in the following: 26 • encountered task constructs 27 • encountered taskyield constructs 28 • encountered taskwait constructs 29 • encountered barrier directives 30 • implicit barrier regions 31 • at the end of the tied task region task switching 32 8 The act of a thread switching from the execution of one task to another task. OpenMP API • Version 3.1 July 2011 1 2 tied task 3 4 untied task A task that, when its task region is suspended, can be resumed by any thread in the team. That is, the task is not tied to any thread. 5 6 7 undeferred task A task for which execution is not deferred with respect to its generating task region. That is, its generating task region is suspended until execution of the undeferred task is completed. 8 9 10 included task A task for which execution is sequentially included in the generating task region. That is, it is undeferred and executed immediately by the encountering thread. 11 12 merged task A task whose data environment, inclusive of ICVs, is the same as that of its generating task region. 13 final task 14 task synchronization construct 15 16 17 1.2.4 variable 27 28 29 A taskwait or a barrier construct. A named data storage block, whose value can be defined and redefined during the execution of a program. Array sections and substrings are not considered variables. private variable With respect to a given set of task regions that bind to the same parallel region, a variable whose name provides access to a different block of storage for each task region. A variable that is part of another variable (as an array or structure element) cannot be made private independently of other components. 22 23 24 25 26 A task that forces all of its child tasks to become final and included tasks. Data Terminology 18 19 20 21 A task that, when its task region is suspended, can be resumed only by the same thread that suspended it. That is, the task is tied to that thread. shared variable With respect to a given set of task regions that bind to the same parallel region, a variable whose name provides access to the same block of storage for each task region. A variable that is part of another variable (as an array or structure element) cannot be shared independently of the other components, except for static data members of C++ classes. Chapter 1 Introduction 9 threadprivate variable 1 2 3 A variable that is replicated, one instance per thread, by the OpenMP implementation. Its name then provides access to a different block of storage for each thread. A variable that is part of another variable (as an array or structure element) cannot be made threadprivate independently of the other components, except for static data members of C++ classes. 4 5 6 threadprivate memory 7 8 9 10 data environment 11 defined The set of threadprivate variables associated with each thread. All the variables associated with the execution of a given task. The data environment for a given task is constructed from the data environment of the generating task at the time the task is generated. For variables, the property of having a valid value. 12 For C: 13 For the contents of variables, the property of having a valid value. 14 For C++: 15 16 For the contents of variables of POD (plain old data) type, the property of having a valid value. 17 18 For variables of non-POD class type, the property of having been constructed but not subsequently destructed. 19 For Fortran: 20 21 22 For the contents of variables, the property of having a valid value. For the allocation or association status of variables, the property of having a valid status. 23 24 COMMENT: Programs that rely upon variables that are not defined are nonconforming programs. class type 25 26 27 28 1.2.5 Implementation Terminology supporting n levels of parallelism 10 For C++: Variables declared with one of the class, struct, or union keywords. Implies allowing an active parallel region to be enclosed by n-1 active parallel regions. OpenMP API • Version 3.1 July 2011 1 supporting the OpenMP API 2 supporting nested parallelism 3 4 internal control variable compliant implementation A conceptual variable that specifies run-time behavior of a set of threads or tasks in an OpenMP program. An implementation of the OpenMP specification that compiles and executes any conforming program as defined by the specification. COMMENT: A compliant implementation may exhibit unspecified behavior when compiling or executing a non-conforming program. 9 10 11 12 Supporting more than one level of parallelism. COMMENT: The acronym ICV is used interchangeably with the term internal control variable in the remainder of this specification. 5 6 7 8 Supporting at least one level of parallelism. unspecified behavior A behavior or result that is not specified by the OpenMP specification or not known prior to the compilation or execution of an OpenMP program. 13 Such unspecified behavior may result from: 14 15 • Issues documented by the OpenMP specification as having unspecified behavior. 16 • A non-conforming program. 17 • A conforming program exhibiting an implementation defined behavior. 18 19 20 21 22 23 implementation defined Behavior that must be documented by the implementation, and is allowed to vary among different compliant implementations. An implementation is allowed to define this behavior as unspecified. COMMENT: All features that have implementation defined behavior are documented in Appendix E. Chapter 1 Introduction 11 1 1.3 Execution Model 2 3 4 5 6 7 8 9 10 11 12 13 The OpenMP API uses the fork-join model of parallel execution. Multiple threads of execution perform tasks defined implicitly or explicitly by OpenMP directives. The OpenMP API is intended to support programs that will execute correctly both as parallel programs (multiple threads of execution and a full OpenMP support library) and as sequential programs (directives ignored and a simple OpenMP stubs library). However, it is possible and permitted to develop a program that executes correctly as a parallel program but not as a sequential program, or that produces different results when executed as a parallel program compared to when it is executed as a sequential program. Furthermore, using different numbers of threads may result in different numeric results because of changes in the association of numeric operations. For example, a serial addition reduction may have a different pattern of addition associations than a parallel reduction. These different associations may change the results of floating-point addition. 14 15 16 17 An OpenMP program begins as a single thread of execution, called the initial thread. The initial thread executes sequentially, as if enclosed in an implicit task region, called the initial task region, that is defined by an implicit inactive parallel region surrounding the whole program. 18 19 20 21 22 23 24 25 26 27 28 When any thread encounters a parallel construct, the thread creates a team of itself and zero or more additional threads and becomes the master of the new team. A set of implicit tasks, one per thread, is generated. The code for each task is defined by the code inside the parallel construct. Each task is assigned to a different thread in the team and becomes tied; that is, it is always executed by the thread to which it is initially assigned. The task region of the task being executed by the encountering thread is suspended, and each member of the new team executes its implicit task. There is an implicit barrier at the end of the parallel construct. Only the master thread resumes execution beyond the end of the parallel construct, resuming the task region that was suspended upon encountering the parallel construct. Any number of parallel constructs can be specified in a single program. 29 30 31 32 33 parallel regions may be arbitrarily nested inside each other. If nested parallelism is disabled, or is not supported by the OpenMP implementation, then the new team that is created by a thread encountering a parallel construct inside a parallel region will consist only of the encountering thread. However, if nested parallelism is supported and enabled, then the new team can consist of more than one thread. 34 35 36 37 38 When any team encounters a worksharing construct, the work inside the construct is divided among the members of the team, and executed cooperatively instead of being executed by every thread. There is a default barrier at the end of each worksharing construct unless the nowait clause is present. Redundant execution of code by every thread in the team resumes after the end of the worksharing construct. 12 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 When any thread encounters a task construct, a new explicit task is generated. Execution of explicitly generated tasks is assigned to one of the threads in the current team, subject to the thread's availability to execute work. Thus, execution of the new task could be immediate, or deferred until later. Threads are allowed to suspend the current task region at a task scheduling point in order to execute a different task. If the suspended task region is for a tied task, the initially assigned thread later resumes execution of the suspended task region. If the suspended task region is for an untied task, then any thread may resume its execution. Completion of all explicit tasks bound to a given parallel region is guaranteed before the master thread leaves the implicit barrier at the end of the region. Completion of a subset of all explicit tasks bound to a given parallel region may be specified through the use of task synchronization constructs. Completion of all explicit tasks bound to the implicit parallel region is guaranteed by the time the program exits. 14 15 16 17 Synchronization constructs and library routines are available in the OpenMP API to coordinate tasks and data access in parallel regions. In addition, library routines and environment variables are available to control or to query the runtime environment of OpenMP programs. 18 19 20 21 22 The OpenMP specification makes no guarantee that input or output to the same file is synchronous when executed in parallel. In this case, the programmer is responsible for synchronizing input and output statements (or routines) using the provided synchronization constructs or library routines. For the case where each thread accesses a different file, no synchronization by the programmer is necessary. 23 24 1.4 Memory Model 25 1.4.1 Structure of the OpenMP Memory Model 26 27 28 29 30 31 32 33 34 35 The OpenMP API provides a relaxed-consistency, shared-memory model. All OpenMP threads have access to a place to store and to retrieve variables, called the memory. In addition, each thread is allowed to have its own temporary view of the memory. The temporary view of memory for each thread is not a required part of the OpenMP memory model, but can represent any kind of intervening structure, such as machine registers, cache, or other local storage, between the thread and the memory. The temporary view of memory allows the thread to cache variables and thereby to avoid going to memory for every reference to a variable. Each thread also has access to another type of memory that must not be accessed by other threads, called threadprivate memory. Chapter 1 Introduction 13 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 A directive that accepts data-sharing attribute clauses determines two kinds of access to variables used in the directive’s associated structured block: shared and private. Each variable referenced in the structured block has an original variable, which is the variable by the same name that exists in the program immediately outside the construct. Each reference to a shared variable in the structured block becomes a reference to the original variable. For each private variable referenced in the structured block, a new version of the original variable (of the same type and size) is created in memory for each task that contains code associated with the directive. Creation of the new version does not alter the value of the original variable. However, the impact of attempts to access the original variable during the region associated with the directive is unspecified; see Section 2.9.3.3 on page 96 for additional details. References to a private variable in the structured block refer to the current task’s private version of the original variable. The relationship between the value of the original variable and the initial or final value of the private version depends on the exact clause that specifies it. Details of this issue, as well as other issues with privatization, are provided in Section 2.9 on page 84. 16 17 18 The minimum size at which a memory update may also read and write back adjacent variables that are part of another variable (as array or structure elements) is implementation defined but is no larger than required by the base language. 19 20 21 22 23 24 A single access to a variable may be implemented with multiple load or store instructions, and hence is not guaranteed to be atomic with respect to other accesses to the same variable. Accesses to variables smaller than the implementation defined minimum size or to C or C++ bit-fields may be implemented by reading, modifying, and rewriting a larger unit of memory, and may thus interfere with updates of variables or fields in the same unit of memory. 25 26 27 28 29 30 If multiple threads write without synchronization to the same memory unit, including cases due to atomicity considerations as described above, then a data race occurs. Similarly, if at least one thread reads from a memory unit and at least one thread writes without synchronization to that same memory unit, including cases due to atomicity considerations as described above, then a data race occurs. If a data race occurs then the result of the program is unspecified. 31 32 33 34 35 36 37 A private variable in a task region that eventually generates an inner nested parallel region is permitted to be made shared by implicit tasks in the inner parallel region. A private variable in a task region can be shared by an explicit task region generated during its execution. However, it is the programmer’s responsibility to ensure through synchronization that the lifetime of the variable does not end before completion of the explicit task region sharing it. Any other access by one task to the private variables of another task results in unspecified behavior. 14 OpenMP API • Version 3.1 July 2011 1 1.4.2 The Flush Operation 2 3 4 5 6 7 The memory model has relaxed-consistency because a thread’s temporary view of memory is not required to be consistent with memory at all times. A value written to a variable can remain in the thread’s temporary view until it is forced to memory at a later time. Likewise, a read from a variable may retrieve the value from the thread’s temporary view, unless it is forced to read from memory. The OpenMP flush operation enforces consistency between the temporary view and memory. 8 9 10 11 12 The flush operation is applied to a set of variables called the flush-set. The flush operation restricts reordering of memory operations that an implementation might otherwise do. Implementations must not reorder the code for a memory operation for a given variable, or the code for a flush operation for the variable, with respect to a flush operation that refers to the same variable. 13 14 15 16 17 18 19 20 21 22 23 24 25 26 If a thread has performed a write to its temporary view of a shared variable since its last flush of that variable, then when it executes another flush of the variable, the flush does not complete until the value of the variable has been written to the variable in memory. If a thread performs multiple writes to the same variable between two flushes of that variable, the flush ensures that the value of the last write is written to the variable in memory. A flush of a variable executed by a thread also causes its temporary view of the variable to be discarded, so that if its next memory operation for that variable is a read, then the thread will read from memory when it may again capture the value in the temporary view. When a thread executes a flush, no later memory operation by that thread for a variable involved in that flush is allowed to start until the flush completes. The completion of a flush of a set of variables executed by a thread is defined as the point at which all writes to those variables performed by the thread before the flush are visible in memory to all other threads and that thread’s temporary view of all variables involved is discarded. 27 28 29 30 31 32 The flush operation provides a guarantee of consistency between a thread’s temporary view and memory. Therefore, the flush operation can be used to guarantee that a value written to a variable by one thread may be read by a second thread. To accomplish this, the programmer must ensure that the second thread has not written to the variable since its last flush of the variable, and that the following sequence of events happens in the specified order: 33 1. The value is written to the variable by the first thread. 34 2. The variable is flushed by the first thread. 35 3. The variable is flushed by the second thread. 36 4. The value is read from the variable by the second thread. Chapter 1 Introduction 15 Note – OpenMP synchronization operations, described in Section 2.8 on page 67 and in Section 3.3 on page 141, are recommended for enforcing this order. Synchronization through variables is possible but is not recommended because the proper timing of flushes is difficult as shown in Section A.2 on page 162. 1 2 3 4 5 1.4.3 OpenMP Memory Consistency The restrictions in Section 1.4.2 on page 15 on reordering with respect to flush operations guarantee the following: 6 7 8 9 10 • If the intersection of the flush-sets of two flushes performed by two different threads 11 12 13 • If two operations performed by the same thread either access, modify, or flush the 14 15 • If the intersection of the flush-sets of two flushes is empty, the threads can observe 16 17 18 The flush operation can be specified using the flush directive, and is also implied at various locations in an OpenMP program: see Section 2.8.6 on page 78 for details. For an example illustrating the memory model, see Section A.2 on page 162. 19 20 Note – Since flush operations by themselves cannot prevent data races, explicit flush operations are only useful in combination with atomic directives. 21 OpenMP programs that: 22 • do not use atomic directives, 23 24 • do not rely on the accuracy of a false result from omp_test_lock and 25 • correctly avoid data races as required in Section 1.4.1 on page 13 26 27 28 29 behave as though operations on shared variables were simply interleaved in an order consistent with the order in which they are performed by each thread. The relaxed consistency model is invisible for such programs, and any explicit flush operations in such programs are redundant. is non-empty, then the two flushes must be completed as if in some sequential order, seen by all threads. same variable, then they must be completed as if in that thread's program order, as seen by all threads. these flushes in any order. omp_test_nest_lock, and 16 OpenMP API • Version 3.1 July 2011 Implementations are allowed to relax the ordering imposed by implicit flush operations when the result is only visible to programs using atomic directives. 1 2 3 1.5 OpenMP Compliance 4 5 6 7 8 An implementation of the OpenMP API is compliant if and only if it compiles and executes all conforming programs according to the syntax and semantics laid out in Chapters 1, 2, 3 and 4. Appendices A, B, C, D, E and F and sections designated as Notes (see Section 1.7 on page 18) are for information purposes only and are not part of the specification. 9 10 11 12 13 14 The OpenMP API defines constructs that operate in the context of the base language that is supported by an implementation. If the base language does not support a language construct that appears in this document, a compliant OpenMP implementation is not required to support it, with the exception that for Fortran, the implementation must allow case insensitivity for directive and API routines names, and must allow identifiers of more than six characters. 15 16 17 18 19 20 All library, intrinsic and built-in routines provided by the base language must be threadsafe in a compliant implementation. In addition, the implementation of the base language must also be thread-safe. For example, ALLOCATE and DEALLOCATE statements must be thread-safe in Fortran. Unsynchronized concurrent use of such routines by different threads must produce correct results (although not necessarily the same as serial execution results, as in the case of random number generation routines). 21 22 23 24 In both Fortran 90 and Fortran 95, variables with explicit initialization have the SAVE attribute implicitly. This is not the case in Fortran 77. However, a compliant OpenMP Fortran implementation must give such a variable the SAVE attribute, regardless of the underlying base language version. 25 26 27 Appendix E lists certain aspects of the OpenMP API that are implementation defined. A compliant implementation is required to define and document its behavior for each of the items in Appendix E. 28 1.6 Normative References 29 30 • ISO/IEC 9899:1990, Information Technology - Programming Languages - C. 31 This OpenMP API specification refers to ISO/IEC 9899:1990 as C90. Chapter 1 Introduction 17 1 2 • ISO/IEC 9899:1999, Information Technology - Programming Languages - C. 3 This OpenMP API specification refers to ISO/IEC 9899:1999 as C99. 4 5 • ISO/IEC 14882:1998, Information Technology - Programming Languages - C++. 6 This OpenMP API specification refers to ISO/IEC 14882:1998 as C++. 7 8 • ISO/IEC 1539:1980, Information Technology - Programming Languages - Fortran. 9 This OpenMP API specification refers to ISO/IEC 1539:1980 as Fortran 77. 10 11 • ISO/IEC 1539:1991, Information Technology - Programming Languages - Fortran. 12 This OpenMP API specification refers to ISO/IEC 1539:1991 as Fortran 90. 13 14 • ISO/IEC 1539-1:1997, Information Technology - Programming Languages - Fortran. 15 This OpenMP API specification refers to ISO/IEC 1539-1:1997 as Fortran 95. 16 Where this OpenMP API specification refers to C, C++ or Fortran, reference is made to the base language supported by the implementation. 17 18 19 1.7 Organization of this document 20 The remainder of this document is structured as follows: 21 • Chapter 2: Directives 22 • Chapter 3: Runtime Library Routines 23 • Chapter 4: Environment Variables 24 • Appendix A: Examples 25 • Appendix B: Stubs for Runtime Library Routines 26 • Appendix C: OpenMP C and C++ Grammar 27 • Appendix D: Interface Declarations 28 • Appendix E: OpenMP Implementation Defined Behaviors 29 • Appendix F: Features History 18 OpenMP API • Version 3.1 July 2011 1 2 3 Some sections of this document only apply to programs written in a certain base language. Text that applies only to programs whose base language is C or C++ is shown as follows: 4 C/C++ specific text.... C/C++ C/C++ 5 Text that applies only to programs whose base language is Fortran is shown as follows: Fortran 6 Fortran specific text...... Fortran 7 8 Where an entire page consists of, for example, Fortran specific text, a marker is shown at the top of the page like this: 9 10 Some text is for information only, and is not part of the normative specification. Such text is designated as a note, like this: 11 Note – Non-normative text.... Fortran (cont.) Chapter 1 Introduction 19 1 This page intentionally left blank. 2 20 OpenMP API • Version 3.1 July 2011 1 CHAPTER 2 2 Directives 3 4 This chapter describes the syntax and behavior of OpenMP directives, and is divided into the following sections: 5 • The language-specific directive format (Section 2.1 on page 22) 6 • Mechanisms to control conditional compilation (Section 2.2 on page 26) 7 • Control of OpenMP API ICVs (Section 2.3 on page 28) 8 9 • Details of each OpenMP directive (Section 2.4 on page 33 to Section 2.10 on page 111) C/C++ 10 11 In C/C++, OpenMP directives are specified by using the #pragma mechanism provided by the C and C++ standards. C/C++ Fortran 12 13 14 In Fortran, OpenMP directives are specified by using special comments that are identified by unique sentinels. Also, a special comment form is available for conditional compilation. Fortran 15 16 17 18 19 20 Compilers can therefore ignore OpenMP directives and conditionally compiled code if support of the OpenMP API is not provided or enabled. A compliant implementation must provide an option or interface that ensures that underlying support of all OpenMP directives and OpenMP conditional compilation mechanisms is enabled. In the remainder of this document, the phrase OpenMP compilation is used to mean a compilation with these OpenMP features enabled. 21 Fortran 1 Restrictions 2 The following restriction applies to all OpenMP directives: 3 • OpenMP directives may not appear in PURE or ELEMENTAL procedures. Fortran 4 2.1 Directive Format C/C++ OpenMP directives for C/C++ are specified with the pragma preprocessing directive. The syntax of an OpenMP directive is formally specified by the grammar in Appendix C, and informally as follows: 5 6 7 #pragma omp directive-name [clause[ [,] clause]...] new-line 8 9 10 11 12 Each directive starts with #pragma omp. The remainder of the directive follows the conventions of the C and C++ standards for compiler directives. In particular, white space can be used before and after the #, and sometimes white space must be used to separate the words in a directive. Preprocessing tokens following the #pragma omp are subject to macro replacement. 13 Directives are case-sensitive. 14 15 An OpenMP executable directive applies to at most one succeeding statement, which must be a structured block. C/C++ Fortran OpenMP directives for Fortran are specified as follows: 16 sentinel directive-name [clause[[,] clause]...] 17 18 19 All OpenMP compiler directives must begin with a directive sentinel. The format of a sentinel differs between fixed and free-form source files, as described in Section 2.1.1 on page 23 and Section 2.1.2 on page 24. 20 21 Directives are case-insensitive. Directives cannot be embedded within continued statements, and statements cannot be embedded within directives. 22 OpenMP API • Version 3.1 July 2011 In order to simplify the presentation, free form is used for the syntax of OpenMP directives for Fortran in the remainder of this document, except as noted. 1 2 Fortran 3 4 5 6 Only one directive-name can be specified per directive (note that this includes combined directives, see Section 2.6 on page 55). The order in which clauses appear on directives is not significant. Clauses on directives may be repeated as needed, subject to the restrictions listed in the description of each clause. 7 8 9 10 Some data-sharing attribute clauses (Section 2.9.3 on page 92), data copying clauses (Section 2.9.4 on page 107), the threadprivate directive (Section 2.9.2 on page 88) and the flush directive (Section 2.8.6 on page 78) accept a list. A list consists of a comma-separated collection of one or more list items. 11 12 A list item is a variable name, subject to the restrictions specified in each of the sections describing clauses and directives for which a list appears. C/C++ C/C++ Fortran A list item is a variable name or a common block name (enclosed in slashes), subject to the restrictions specified in each of the sections describing clauses and directives for which a list appears. 13 14 15 Fortran 16 Fortran 17 18 2.1.1 Fixed Source Form Directives The following sentinels are recognized in fixed form source files: !$omp | c$omp | *$omp 19 20 21 22 23 Sentinels must start in column 1 and appear as a single word with no intervening characters. Fortran fixed form line length, white space, continuation, and column rules apply to the directive line. Initial directive lines must have a space or zero in column 6, and continuation directive lines must have a character other than a space or a zero in column 6. 24 25 26 27 Comments may appear on the same line as a directive. The exclamation point initiates a comment when it appears after column 6. The comment extends to the end of the source line and is ignored. If the first non-blank character after the directive sentinel of an initial or continuation directive line is an exclamation point, the line is ignored. Chapter 2 Directives 23 Fortran (cont.) 1 2 3 Note – in the following example, the three formats for specifying the directive are equivalent (the first line represents the position of the first 9 columns): 4 c23456789 5 !$omp parallel do shared(a,b,c) 6 7 c$omp parallel do 8 c$omp+shared(a,b,c) 9 10 11 c$omp paralleldoshared(a,b,c) 2.1.2 Free Source Form Directives The following sentinel is recognized in free form source files: 12 !$omp 13 14 15 16 17 18 19 20 The sentinel can appear in any column as long as it is preceded only by white space (spaces and tab characters). It must appear as a single word with no intervening character. Fortran free form line length, white space, and continuation rules apply to the directive line. Initial directive lines must have a space after the sentinel. Continued directive lines must have an ampersand (&) as the last nonblank character on the line, prior to any comment placed inside the directive. Continuation directive lines can have an ampersand after the directive sentinel with optional white space before and after the ampersand. 21 22 23 24 Comments may appear on the same line as a directive. The exclamation point (!) initiates a comment. The comment extends to the end of the source line and is ignored. If the first nonblank character after the directive sentinel is an exclamation point, the line is ignored. 25 26 27 One or more blanks or horizontal tabs must be used to separate adjacent keywords in directives in free source form, except in the following cases, where white space is optional between the given pair of keywords: 24 OpenMP API • Version 3.1 July 2011 1 end atomic end critical end do end master end ordered end parallel end sections end single end task end workshare parallel do parallel sections parallel workshare 2 3 Note – in the following example the three formats for specifying the directive are equivalent (the first line represents the position of the first 9 columns): 4 !23456789 5 6 !$omp parallel do & !$omp shared(a,b,c) 7 8 9 !$omp parallel & !$omp&do shared(a,b,c) 10 11 !$omp paralleldo shared(a,b,c) 12 Fortran Chapter 2 Directives 25 1 2.2 Conditional Compilation 2 3 4 In implementations that support a preprocessor, the _OPENMP macro name is defined to have the decimal value yyyymm where yyyy and mm are the year and month designations of the version of the OpenMP API that the implementation supports. 5 6 If this macro is the subject of a #define or a #undef preprocessing directive, the behavior is unspecified. 7 For examples of conditional compilation, see Section A.3 on page 169. Fortran The OpenMP API requires Fortran lines to be compiled conditionally, as described in the following sections. 8 9 11 Fixed Source Form Conditional Compilation Sentinels 12 13 The following conditional compilation sentinels are recognized in fixed form source files: 10 2.2.1 !$ | *$ | c$ 14 15 To enable conditional compilation, a line with a conditional compilation sentinel must satisfy the following criteria: 16 17 • The sentinel must start in column 1 and appear as a single word with no intervening 18 19 • After the sentinel is replaced with two spaces, initial lines must have a space or zero 20 21 22 • After the sentinel is replaced with two spaces, continuation lines must have a 23 24 If these criteria are met, the sentinel is replaced by two spaces. If these criteria are not met, the line is left unchanged. white space. in column 6 and only white space and numbers in columns 1 through 5. character other than a space or zero in column 6 and only white space in columns 1 through 5. 25 26 27 26 OpenMP API • Version 3.1 July 2011 Fortran (cont.) 1 2 3 4 Note – in the following example, the two forms for specifying conditional compilation in fixed source form are equivalent (the first line represents the position of the first 9 columns): 5 c23456789 6 !$ 10 iam = omp_get_thread_num() + 7 !$ & index 8 9 10 #ifdef _OPENMP 11 10 iam = omp_get_thread_num() + 12 & 13 #endif 15 Free Source Form Conditional Compilation Sentinel 16 The following conditional compilation sentinel is recognized in free form source files: 14 2.2.2 index !$ 17 18 To enable conditional compilation, a line with a conditional compilation sentinel must satisfy the following criteria: 19 • The sentinel can appear in any column but must be preceded only by white space. 20 • The sentinel must appear as a single word with no intervening white space. 21 • Initial lines must have a space after the sentinel. 22 23 24 25 • Continued lines must have an ampersand as the last nonblank character on the line, 26 27 If these criteria are met, the sentinel is replaced by two spaces. If these criteria are not met, the line is left unchanged. prior to any comment appearing on the conditionally compiled line. Continued lines can have an ampersand after the sentinel, with optional white space before and after the ampersand. Chapter 2 Directives 27 1 2 3 Note – in the following example, the two forms for specifying conditional compilation in free source form are equivalent (the first line represents the position of the first 9 columns): 4 c23456789 5 !$ iam = omp_get_thread_num() + 6 !$& & index 7 8 #ifdef _OPENMP 9 iam = omp_get_thread_num() + 10 & index 11 #endif 12 Fortran 13 2.3 Internal Control Variables 14 15 16 17 18 19 20 21 An OpenMP implementation must act as if there were internal control variables (ICVs) that control the behavior of an OpenMP program. These ICVs store information such as the number of threads to use for future parallel regions, the schedule to use for worksharing loops and whether nested parallelism is enabled or not. The ICVs are given values at various times (described below) during the execution of the program. They are initialized by the implementation itself and may be given values through OpenMP environment variables and through calls to OpenMP API routines. The program can retrieve the values of these ICVs only through OpenMP API routines. 22 23 24 For purposes of exposition, this document refers to the ICVs by certain names, but an implementation is not required to use these names or to offer any way to access the variables other than through the ways shown in Section 2.3.2 on page 29. 25 2.3.1 ICV Descriptions The following ICVs store values that affect the operation of parallel regions. 26 27 28 OpenMP API • Version 3.1 July 2011 1 2 3 • dyn-var - controls whether dynamic adjustment of the number of threads is enabled 4 5 • nest-var - controls whether nested parallelism is enabled for encountered parallel 6 7 • nthreads-var - controls the number of threads requested for encountered parallel 8 9 • thread-limit-var - controls the maximum number of threads participating in the for encountered parallel regions. There is one copy of this ICV per data environment. regions. There is one copy of this ICV per data environment. regions. There is one copy of this ICV per data environment. OpenMP program. There is one copy of this ICV for the whole program. 10 11 • max-active-levels-var - controls the maximum number of nested active parallel 12 The following ICVs store values that affect the operation of loop regions. 13 14 • run-sched-var - controls the schedule that the runtime schedule clause uses for 15 16 • def-sched-var - controls the implementation defined default scheduling of loop 17 The following ICVs store values that affect the program execution. 18 19 20 • bind-var - controls the binding of threads to processors. If binding is enabled, the 21 22 • stacksize-var - controls the stack size for threads that the OpenMP implementation 23 24 • wait-policy-var - controls the desired behavior of waiting threads. There is one copy 25 regions. There is one copy of this ICV for the whole program. loop regions. There is one copy of this ICV per data environment. regions. There is one copy of this ICV for the whole program. execution environment is advised not to move OpenMP threads between processors. There is one copy of this ICV for the whole program. creates. There is one copy this ICV for the whole program. of this ICV for the whole program. 2.3.2 Modifying and Retrieving ICV Values The following table shows the methods for retrieving the values of the ICVs as well as their initial values: 26 27 ICV Scope Ways to modify value Way to retrieve value Initial value dyn-var data environment OMP_DYNAMIC omp_set_dynamic() omp_get_dynamic() See comments below nest-var data environment OMP_NESTED omp_set_nested() omp_get_nested() false nthreads-var data environment OMP_NUM_THREADS omp_set_num_threads() omp_get_max_threads() Implementation defined run-sched-var data environment OMP_SCHEDULE omp_set_schedule() omp_get_schedule() Implementation defined Chapter 2 Directives 29 ICV Scope Ways to modify value Way to retrieve value Initial value def-sched-var global (none) (none) Implementation defined bind-var global OMP_PROC_BIND (none) Implementation defined stacksize-var global OMP_STACKSIZE (none) Implementation defined wait-policy-var global OMP_WAIT_POLICY (none) Implementation defined thread-limit-var global OMP_THREAD_LIMIT omp_get_thread_limit() Implementation defined max-active-levels-var global OMP_MAX_ACTIVE_LEVELS omp_set_max_active_ levels() omp_get_max_active_ levels() See comments below 1 Comments: 2 3 4 • The value of the nthreads-var ICV is a list. The runtime call 5 6 • The initial value of dyn-var is implementation defined if the implementation supports 7 8 9 • The initial value of max-active-levels-var is the number of levels of parallelism that 10 11 12 13 After the initial values are assigned, but before any OpenMP construct or OpenMP API routine executes, the values of any OpenMP environment variables that were set by the user are read and the associated ICVs are modified accordingly. After this point, no changes to any OpenMP environment variables will affect the ICVs. 14 Clauses on OpenMP constructs do not modify the values of any of the ICVs. 15 omp_set_num_threads() sets the value of the first element of this list, and omp_get_max_threads() retrieves the value of the first element of this list. dynamic adjustment of the number of threads; otherwise, the initial value is false. the implementation supports. See the definition of supporting n levels of parallelism in Section 1.2.5 on page 10 for further details. 2.3.3 How the Per-Data Environment ICVs Work 16 17 Each data environment has its own copies of internal variables dyn-var, nest-var, nthreads-var, and run-sched-var. 18 19 20 Calls to omp_set_num_threads(), omp_set_dynamic(), omp_set_nested(), and omp_set_schedule() modify only the ICVs in the data environment of their binding task. 21 22 23 When a task construct or parallel construct is encountered, the generated task(s) inherit the values of dyn-var, nest-var, and run-sched-var from the generating task's ICV values. 30 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 When a task construct is encountered, the generated task inherits the value of nthreads-var from the generating task's nthreads-var value. When a parallel construct is encountered, and the generating task's nthreads-var list contains a single element, the generated task(s) inherit that list as the value of nthreads-var. When a parallel construct is encountered, and the generating task's nthreads-var list contains multiple elements, the generated task(s) inherit the value of nthreads-var as the list obtained by deletion of the first element from the generating task's nthreads-var value. 8 9 10 When encountering a loop worksharing region with schedule(runtime), all implicit task regions that constitute the binding parallel region must have the same value for run-sched-var in their data environments. Otherwise, the behavior is unspecified. 11 12 13 2.3.4 ICV Override Relationships The override relationships among various construct clauses, OpenMP API routines, environment variables, and the initial values of ICVs are shown in the following table: 14 construct clause, if used overrides call to API routine overrides setting of environment variable overrides initial value of (none) (none) omp_set_dynamic() OMP_DYNAMIC dyn-var omp_set_nested() OMP_NESTED nest-var num_threads omp_set_num_threads() OMP_NUM_THREADS nthreads-var schedule omp_set_schedule() OMP_SCHEDULE run-sched-var (none) (none) OMP_PROC_BIND bind-var schedule (none) (none) def-sched-var (none) (none) OMP_STACKSIZE stacksize-var (none) (none) OMP_WAIT_POLICY wait-policy-var (none) (none) OMP_THREAD_LIMIT thread-limit-var (none) omp_set_max_active_levels() OMP_MAX_ACTIVE_LEVELS max-active-levels-var * 15 16 17 * The num_threads clause and omp_set_num_threads() override the value of the OMP_NUM_THREADS environment variable and the initial value of the first element of the nthreads-var ICV. 18 Cross References: 19 • parallel construct, see Section 2.4 on page 33. 20 • num_threads clause, see Section 2.4.1 on page 36. Chapter 2 Directives 31 1 • schedule clause, see Section 2.5.1.1 on page 47. 2 • Loop construct, see Section 2.5.1 on page 39. 3 • omp_set_num_threads routine, see Section 3.2.1 on page 116. 4 • omp_get_max_threads routine, see Section 3.2.3 on page 118. 5 • omp_set_dynamic routine, see Section 3.2.7 on page 123. 6 • omp_get_dynamic routine, see Section 3.2.8 on page 124. 7 • omp_set_nested routine, see Section 3.2.9 on page 125. 8 • omp_get_nested routine, see Section 3.2.10 on page 126. 9 • omp_set_schedule routine, see Section 3.2.11 on page 128. 10 • omp_get_schedule routine, see Section 3.2.12 on page 130. 11 • omp_get_thread_limit routine, see Section 3.2.13 on page 131. 12 • omp_set_max_active_levels routine, see Section 3.2.14 on page 132. 13 • omp_get_max_active_levels routine, see Section 3.2.15 on page 134. 14 • OMP_SCHEDULE environment variable, see Section 4.1 on page 154. 15 • OMP_NUM_THREADS environment variable, see Section 4.2 on page 155. 16 • OMP_DYNAMIC environment variable, see Section 4.3 on page 156. 17 • OMP_PROC_BIND environment variable, see Section 4.4 on page 156 18 • OMP_NESTED environment variable, see Section 4.5 on page 157. 19 • OMP_STACKSIZE environment variable, see Section 4.6 on page 157. 20 • OMP_WAIT_POLICY environment variable, see Section 4.7 on page 158. 21 • OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.8 on page 159. 22 • OMP_THREAD_LIMIT environment variable, see Section 4.9 on page 160. 32 OpenMP API • Version 3.1 July 2011 1 2.4 parallel Construct 2 Summary 3 4 This fundamental construct starts parallel execution. See Section 1.3 on page 12 for a general description of the OpenMP execution model. 5 Syntax 6 The syntax of the parallel construct is as follows: C/C++ #pragma omp parallel [clause[ [, ]clause] ...] new-line structured-block 7 where clause is one of the following: if(scalar-expression) num_threads(integer-expression) default(shared | none) private(list) firstprivate(list) shared(list) copyin(list) reduction(operator: list) C/C++ 8 Fortran 9 The syntax of the parallel construct is as follows: !$omp parallel [clause[[,] clause]...] structured-block !$omp end parallel Chapter 2 Directives 33 where clause is one of the following: 1 if(scalar-logical-expression) num_threads(scalar-integer-expression) default(private | firstprivate | shared | none) private(list) firstprivate(list) shared(list) copyin(list) reduction({operator|intrinsic_procedure_name}:list) The end parallel directive denotes the end of the parallel construct. 2 Fortran 3 Binding 4 5 The binding thread set for a parallel region is the encountering thread. The encountering thread becomes the master thread of the new team. 6 Description 7 8 9 10 11 12 13 14 When a thread encounters a parallel construct, a team of threads is created to execute the parallel region (see Section 2.4.1 on page 36 for more information about how the number of threads in the team is determined, including the evaluation of the if and num_threads clauses). The thread that encountered the parallel construct becomes the master thread of the new team, with a thread number of zero for the duration of the new parallel region. All threads in the new team, including the master thread, execute the region. Once the team is created, the number of threads in the team remains constant for the duration of that parallel region. 15 16 17 18 Within a parallel region, thread numbers uniquely identify each thread. Thread numbers are consecutive whole numbers ranging from zero for the master thread up to one less than the number of threads in the team. A thread may obtain its own thread number by a call to the omp_get_thread_num library routine. 19 20 21 22 A set of implicit tasks, equal in number to the number of threads in the team, is generated by the encountering thread. The structured block of the parallel construct determines the code that will be executed in each implicit task. Each task is assigned to a different thread in the team and becomes tied. The task region of the task being 34 OpenMP API • Version 3.1 July 2011 1 2 3 executed by the encountering thread is suspended and each thread in the team executes its implicit task. Each thread can execute a path of statements that is different from that of the other threads. 4 5 6 7 The implementation may cause any thread to suspend execution of its implicit task at a task scheduling point, and switch to execute any explicit task generated by any of the threads in the team, before eventually resuming execution of the implicit task (for more details see Section 2.7 on page 61). 8 9 10 There is an implied barrier at the end of a parallel region. After the end of a parallel region, only the master thread of the team resumes execution of the enclosing task region. 11 12 13 If a thread in a team executing a parallel region encounters another parallel directive, it creates a new team, according to the rules in Section 2.4.1 on page 36, and it becomes the master of that new team. 14 15 16 17 18 If execution of a thread terminates while inside a parallel region, execution of all threads in all teams terminates. The order of termination of threads is unspecified. All work done by a team prior to any barrier that the team has passed in the program is guaranteed to be complete. The amount of work done by each thread after the last barrier that it passed and before it terminates is unspecified. 19 20 For an example of the parallel construct, see Section A.5 on page 172. For an example of the num_threads clause, see Section A.7 on page 177. 21 Restrictions 22 Restrictions to the parallel construct are as follows: 23 • A program that branches into or out of a parallel region is non-conforming. 24 25 • A program must not depend on any ordering of the evaluations of the clauses of the 26 • At most one if clause can appear on the directive. 27 28 • At most one num_threads clause can appear on the directive. The num_threads 29 30 31 • A throw executed inside a parallel region must cause execution to resume parallel directive, or on any side effects of the evaluations of the clauses. expression must evaluate to a positive integer value. C/C++ within the same parallel region, and the same thread that threw the exception must catch it. C/C++ Chapter 2 Directives 35 Fortran • Unsynchronized use of Fortran I/O statements by multiple threads on the same unit 1 2 has unspecified behavior. Fortran 3 Cross References 4 5 • default, shared, private, firstprivate, and reduction clauses, see 6 • copyin clause, see Section 2.9.4 on page 107. 7 • omp_get_thread_num routine, see Section 3.2.4 on page 119. 8 Section 2.9.3 on page 92. 2.4.1 9 Determining the Number of Threads for a parallel Region 10 11 12 13 When execution encounters a parallel directive, the value of the if clause or num_threads clause (if any) on the directive, the current parallel context, and the values of the nthreads-var, dyn-var, thread-limit-var, max-active-level-var, and nest-var ICVs are used to determine the number of threads to use in the region. 14 15 16 17 18 19 Note that using a variable in an if or num_threads clause expression of a parallel construct causes an implicit reference to the variable in all enclosing constructs. The if clause expression and the num_threads clause expression are evaluated in the context outside of the parallel construct, and no ordering of those evaluations is specified. It is also unspecified whether, in what order, or how many times any side-effects of the evaluation of the num_threads or if clause expressions occur. 20 21 When a thread encounters a parallel construct, the number of threads is determined according to Algorithm 2.1. Algorithm 2.1 let ThreadsBusy be the number of OpenMP threads currently executing; let ActiveParRegions be the number of enclosing active parallel regions; if an if clause exists then let IfClauseValue be the value of the if clause expression; else let IfClauseValue = true; if a num_threads clause exists 36 OpenMP API • Version 3.1 July 2011 Algorithm 2.1 then let ThreadsRequested be the value of the num_threads clause expression; else let ThreadsRequested = value of the first element of nthreads-var; let ThreadsAvailable = (thread-limit-var - ThreadsBusy + 1); if (IfClauseValue = false) then number of threads = 1; else if (ActiveParRegions >= 1) and (nest-var = false) then number of threads = 1; else if (ActiveParRegions = max-active-levels-var) then number of threads = 1; else if (dyn-var = true) and (ThreadsRequested <= ThreadsAvailable) then number of threads = [ 1 : ThreadsRequested ]; else if (dyn-var = true) and (ThreadsRequested > ThreadsAvailable) then number of threads = [ 1 : ThreadsAvailable ]; else if (dyn-var = false) and (ThreadsRequested <= ThreadsAvailable) then number of threads = ThreadsRequested; else if (dyn-var = false) and (ThreadsRequested > ThreadsAvailable) then behavior is implementation defined; 1 2 3 Note – Since the initial value of the dyn-var ICV is implementation defined, programs that depend on a specific number of threads for correct execution should explicitly disable dynamic adjustment of the number of threads. 4 Cross References 5 6 • nthreads-var, dyn-var, thread-limit-var, max-active-level-var, and nest-var ICVs, see Section 2.3 on page 28. Chapter 2 Directives 37 1 2.5 Worksharing Constructs A worksharing construct distributes the execution of the associated region among the members of the team that encounters it. Threads execute portions of the region in the context of the implicit tasks each one is executing. If the team consists of only one thread then the worksharing region is not executed in parallel. 2 3 4 5 6 7 8 9 10 11 12 A worksharing region has no barrier on entry; however, an implied barrier exists at the end of the worksharing region, unless a nowait clause is specified. If a nowait clause is present, an implementation may omit the barrier at the end of the worksharing region. In this case, threads that finish early may proceed straight to the instructions following the worksharing region without waiting for the other members of the team to finish the worksharing region, and without performing a flush operation (see Section A.10 on page 182 for an example). 13 14 The OpenMP API defines the following worksharing constructs, and these are described in the sections that follow: 15 • loop construct 16 • sections construct 17 • single construct 18 • workshare construct 19 Restrictions 20 The following restrictions apply to worksharing constructs: 21 22 • Each worksharing region must be encountered by all threads in a team or by none at 23 24 • The sequence of worksharing regions and barrier regions encountered must be the all. same for every thread in a team. 38 OpenMP API • Version 3.1 July 2011 1 2.5.1 Loop Construct 2 Summary 3 4 5 6 The loop construct specifies that the iterations of one or more associated loops will be executed in parallel by threads in the team in the context of their implicit tasks. The iterations are distributed across threads that already exist in the team executing the parallel region to which the loop region binds. 7 Syntax 8 The syntax of the loop construct is as follows: C/C++ #pragma omp for [clause[[,] clause] ... ] new-line for-loops 9 where clause is one of the following: private(list) firstprivate(list) lastprivate(list) reduction(operator: list) schedule(kind[, chunk_size]) collapse(n) ordered nowait Chapter 2 Directives 39 C/C++ (cont.) The for directive places restrictions on the structure of all associated for-loops. Specifically, all associated for-loops must have the following canonical form: 1 2 for (init-expr; test-expr; incr-expr) structured-block 40 init-expr One of the following: var = lb integer-type var = lb random-access-iterator-type var = lb pointer-type var = lb test-expr One of the following: var relational-op b b relational-op var incr-expr One of the following: ++var var++ --var var-var += incr var -= incr var = var + incr var = incr + var var = var - incr var One of the following: A variable of a signed or unsigned integer type. For C++, a variable of a random access iterator type. For C, a variable of a pointer type. If this variable would otherwise be shared, it is implicitly made private in the loop construct. This variable must not be modified during the execution of the for-loop other than in incr-expr. Unless the variable is specified lastprivate on the loop construct, its value after the loop is unspecified. relational-op One of the following: < <= > >= lb and b Loop invariant expressions of a type compatible with the type of var. incr A loop invariant integer expression. OpenMP API • Version 3.1 July 2011 1 2 3 The canonical form allows the iteration count of all associated loops to be computed before executing the outermost loop. The computation is performed for each loop in an integer type. This type is derived from the type of var as follows: 4 • If var is of an integer type, then the type is the type of var. 5 6 • For C++, if var is of a random access iterator type, then the type is the type that 7 • For C, if var is of a pointer type, then the type is ptrdiff_t. 8 9 The behavior is unspecified if any intermediate result required to compute the iteration count cannot be represented in the type determined above. 10 11 12 There is no implied synchronization during the evaluation of the lb, b, or incr expressions. It is unspecified whether, in what order, or how many times any side effects within the lb, b, or incr expressions occur. 13 14 15 16 Note – Random access iterators are required to support random access to elements in constant time. Other iterators are precluded by the restrictions since they can take linear time or offer limited functionality. It is therefore advisable to use tasks to parallelize those cases. would be used by std::distance applied to variables of the type of var. C/C++ 17 Fortran 18 The syntax of the loop construct is as follows: !$omp do [clause[[,] clause] ... ] do-loops [!$omp end do [nowait] ] 19 where clause is one of the following: private(list) firstprivate(list) lastprivate(list) reduction({operator|intrinsic_procedure_name}:list) Chapter 2 Directives 41 schedule(kind[, chunk_size]) collapse(n) ordered 1 2 If an end do directive is not specified, an end do directive is assumed at the end of the do-loop. 3 4 5 6 All associated do-loops must be do-constructs as defined by the Fortran standard. If an end do directive follows a do-construct in which several loop statements share a DO termination statement, then the directive can only be specified for the outermost of these DO statements. See Section A.8 on page 179 for examples. 7 8 9 10 If any of the loop iteration variables would otherwise be shared, they are implicitly made private on the loop construct. See Section A.9 on page 181 for examples. Unless the loop iteration variables are specified lastprivate on the loop construct, their values after the loop are unspecified. Fortran 11 Binding 12 13 14 15 The binding thread set for a loop region is the current team. A loop region binds to the innermost enclosing parallel region. Only the threads of the team executing the binding parallel region participate in the execution of the loop iterations and the implied barrier of the loop region if the barrier is not eliminated by a nowait clause. 16 Description 17 18 The loop construct is associated with a loop nest consisting of one or more loops that follow the directive. 19 20 There is an implicit barrier at the end of a loop construct unless a nowait clause is specified. 21 22 23 24 The collapse clause may be used to specify how many loops are associated with the loop construct. The parameter of the collapse clause must be a constant positive integer expression. If no collapse clause is present, the only loop that is associated with the loop construct is the one that immediately follows the loop directive. 25 26 27 28 If more than one loop is associated with the loop construct, then the iterations of all associated loops are collapsed into one larger iteration space that is then divided according to the schedule clause. The sequential execution of the iterations in all associated loops determines the order of the iterations in the collapsed iteration space. 42 OpenMP API • Version 3.1 July 2011 1 2 3 The iteration count for each associated loop is computed before entry to the outermost loop. If execution of any associated loop changes any of the values used to compute any of the iteration counts, then the behavior is unspecified. 4 5 The integer type (or kind, for Fortran) used to compute the iteration count for the collapsed loop is implementation defined. 6 7 8 9 10 11 12 13 14 15 16 A worksharing loop has logical iterations numbered 0,1,...,N-1 where N is the number of loop iterations, and the logical numbering denotes the sequence in which the iterations would be executed if the associated loop(s) were executed by a single thread. The schedule clause specifies how iterations of the associated loops are divided into contiguous non-empty subsets, called chunks, and how these chunks are distributed among threads of the team. Each thread executes its assigned chunk(s) in the context of its implicit task. The chunk_size expression is evaluated using the original list items of any variables that are made private in the loop construct. It is unspecified whether, in what order, or how many times, any side-effects of the evaluation of this expression occur. The use of a variable in a schedule clause expression of a loop construct causes an implicit reference to the variable in all enclosing constructs. 17 18 19 20 21 Different loop regions with the same schedule and iteration count, even if they occur in the same parallel region, can distribute iterations among threads differently. The only exception is for the static schedule as specified in Table 2-1. Programs that depend on which thread executes a particular iteration under any other circumstances are non-conforming. 22 23 See Section 2.5.1.1 on page 47 for details of how the schedule for a worksharing loop is determined. 24 The schedule kind can be one of those specified in Table 2-1. 25 Chapter 2 Directives 43 1 TABLE 2-1 static schedule clause kind values When schedule(static, chunk_size) is specified, iterations are divided into chunks of size chunk_size, and the chunks are assigned to the threads in the team in a round-robin fashion in the order of the thread number. When no chunk_size is specified, the iteration space is divided into chunks that are approximately equal in size, and at most one chunk is distributed to each thread. Note that the size of the chunks is unspecified in this case. A compliant implementation of the static schedule must ensure that the same assignment of logical iteration numbers to threads will be used in two loop regions if the following conditions are satisfied: 1) both loop regions have the same number of loop iterations, 2) both loop regions have the same value of chunk_size specified, or both loop regions have no chunk_size specified, and 3) both loop regions bind to the same parallel region. A data dependence between the same logical iterations in two such loops is guaranteed to be satisfied allowing safe use of the nowait clause (see Section A.10 on page 182 for examples). dynamic When schedule(dynamic, chunk_size) is specified, the iterations are distributed to threads in the team in chunks as the threads request them. Each thread executes a chunk of iterations, then requests another chunk, until no chunks remain to be distributed. Each chunk contains chunk_size iterations, except for the last chunk to be distributed, which may have fewer iterations. When no chunk_size is specified, it defaults to 1. guided When schedule(guided, chunk_size) is specified, the iterations are assigned to threads in the team in chunks as the executing threads request them. Each thread executes a chunk of iterations, then requests another chunk, until no chunks remain to be assigned. For a chunk_size of 1, the size of each chunk is proportional to the number of unassigned iterations divided by the number of threads in the team, decreasing to 1. For a chunk_size with value k (greater than 1), the size of each chunk is determined in the same way, with the restriction that the chunks do not contain fewer than k iterations (except for the last chunk to be assigned, which may have fewer than k iterations). When no chunk_size is specified, it defaults to 1. auto 44 When schedule(auto) is specified, the decision regarding scheduling is delegated to the compiler and/or runtime system. The programmer gives the implementation the freedom to choose any possible mapping of iterations to threads in the team. OpenMP API • Version 3.1 July 2011 runtime When schedule(runtime) is specified, the decision regarding scheduling is deferred until run time, and the schedule and chunk size are taken from the run-sched-var ICV. If the ICV is set to auto, the schedule is implementation defined. 1 2 3 4 5 6 Note – For a team of p threads and a loop of n iterations, let n ⁄ p be the integer q that satisfies n = p*q - r, with 0 ≤ r < p . One compliant implementation of the static schedule (with no specified chunk_size) would behave as though chunk_size had been specified with value q. Another compliant implementation would assign q iterations to the first p-r threads, and q-1 iterations to the remaining r threads. This illustrates why a conforming program must not rely on the details of a particular implementation. 7 8 9 10 11 12 A compliant implementation of the guided schedule with a chunk_size value of k would assign q = n ⁄ p iterations to the first available thread and set n to the larger of n-q and p*k. It would then repeat this process until q is greater than or equal to the number of remaining iterations, at which time the remaining iterations form the final chunk. Another compliant implementation could use the same method, except with q = n ⁄ ( 2p ) , and set n to the larger of n-q and 2*p*k. 13 Restrictions 14 Restrictions to the loop construct are as follows: 15 16 • All loops associated with the loop construct must be perfectly nested; that is, there 17 18 • The values of the loop control expressions of the loops associated with the loop 19 • Only one schedule clause can appear on a loop directive. 20 • Only one collapse clause can appear on a loop directive. 21 • chunk_size must be a loop invariant integer expression with a positive value. 22 • The value of the chunk_size expression must be the same for all threads in the team. 23 • The value of the run-sched-var ICV must be the same for all threads in the team. 24 25 • When schedule(runtime) or schedule(auto) is specified, chunk_size must 26 • Only one ordered clause can appear on a loop directive. 27 28 • The ordered clause must be present on the loop construct if any ordered region 29 • The loop iteration variable may not appear in a threadprivate directive. must be no intervening code nor any OpenMP directive between any two loops. construct must be the same for all the threads in the team. not be specified. ever binds to a loop region arising from the loop construct. Chapter 2 Directives 45 C/C++ 1 • The associated for-loops must be structured blocks. 2 3 • Only an iteration of the innermost associated loop may be curtailed by a continue 4 • No statement can branch to any associated for statement. 5 • Only one nowait clause can appear on a for directive. 6 7 8 9 • If test-expr is of the form var relational-op b and relational-op is < or <= then 10 11 12 13 • If test-expr is of the form b relational-op var and relational-op is < or <= then 14 15 16 • A throw executed inside a loop region must cause execution to resume within the statement. incr-expr must cause var to increase on each iteration of the loop. If test-expr is of the form var relational-op b and relational-op is > or >= then incr-expr must cause var to decrease on each iteration of the loop. incr-expr must cause var to decrease on each iteration of the loop. If test-expr is of the form b relational-op var and relational-op is > or >= then incr-expr must cause var to increase on each iteration of the loop. same iteration of the loop region, and the same thread that threw the exception must catch it. C/C++ Fortran 17 • The associated do-loops must be structured blocks. 18 19 • Only an iteration of the innermost associated loop may be curtailed by a CYCLE 20 21 • No statement in the associated loops other than the DO statements can cause a branch 22 • The do-loop iteration variable must be of type integer. 23 • The do-loop cannot be a DO WHILE or a DO loop without loop control. statement. out of the loops. Fortran 24 Cross References 25 26 • private, firstprivate, lastprivate, and reduction clauses, see 27 • OMP_SCHEDULE environment variable, see Section 4.1 on page 154. 28 • ordered construct, see Section 2.8.7 on page 82. Section 2.9.3 on page 92. 46 OpenMP API • Version 3.1 July 2011 1 2.5.1.1 Determining the Schedule of a Worksharing Loop 2 3 4 5 6 7 8 9 10 When execution encounters a loop directive, the schedule clause (if any) on the directive, and the run-sched-var and def-sched-var ICVs are used to determine how loop iterations are assigned to threads. See Section 2.3 on page 28 for details of how the values of the ICVs are determined. If the loop directive does not have a schedule clause then the current value of the def-sched-var ICV determines the schedule. If the loop directive has a schedule clause that specifies the runtime schedule kind then the current value of the run-sched-var ICV determines the schedule. Otherwise, the value of the schedule clause determines the schedule. Figure 2-1 describes how the schedule for a worksharing loop is determined. 11 Cross References 12 • ICVs, see Section 2.3 on page 28. 13 START schedule clause present? No Use def-sched-var schedule kind Yes schedule kind value is runtime? No Use schedule kind specified in schedule clause Yes Use run-sched-var schedule kind 14 FIGURE 2-1 Determining the schedule for a worksharing loop. Chapter 2 Directives 47 1 2.5.2 sections Construct 2 Summary 3 4 5 6 The sections construct is a noniterative worksharing construct that contains a set of structured blocks that are to be distributed among and executed by the threads in a team. Each structured block is executed once by one of the threads in the team in the context of its implicit task. 7 Syntax 8 The syntax of the sections construct is as follows: C/C++ #pragma omp sections [clause[[,] clause] ...] new-line { [#pragma omp section new-line] structured-block [#pragma omp section new-line structured-block ] ... } where clause is one of the following: 9 private(list) firstprivate(list) lastprivate(list) reduction(operator: list) nowait 10 C/C++ 48 OpenMP API • Version 3.1 July 2011 Fortran 1 The syntax of the sections construct is as follows: !$omp sections [clause[[,] clause] ...] [!$omp section] structured-block [!$omp section structured-block ] ... !$omp end sections [nowait] 2 where clause is one of the following: private(list) firstprivate(list) lastprivate(list) reduction({operator|intrinsic_procedure_name}:list) 3 Fortran 4 Binding 5 6 7 8 9 The binding thread set for a sections region is the current team. A sections region binds to the innermost enclosing parallel region. Only the threads of the team executing the binding parallel region participate in the execution of the structured blocks and the implied barrier of the sections region if the barrier is not eliminated by a nowait clause. 10 Description 11 12 Each structured block in the sections construct is preceded by a section directive except possibly the first block, for which a preceding section directive is optional. 13 14 The method of scheduling the structured blocks among the threads in the team is implementation defined. 15 16 There is an implicit barrier at the end of a sections construct unless a nowait clause is specified. Chapter 2 Directives 49 1 Restrictions 2 Restrictions to the sections construct are as follows: 3 4 5 • Orphaned section directives are prohibited. That is, the section directives must 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. appear within the sections construct and must not be encountered elsewhere in the sections region. C/C++ • A throw executed inside a sections region must cause execution to resume within 8 9 10 the same section of the sections region, and the same thread that threw the exception must catch it. C/C++ 11 Cross References 12 13 • private, firstprivate, lastprivate, and reduction clauses, see 14 Section 2.9.3 on page 92. 2.5.3 single Construct 15 Summary 16 17 18 19 The single construct specifies that the associated structured block is executed by only one of the threads in the team (not necessarily the master thread), in the context of its implicit task. The other threads in the team, which do not execute the block, wait at an implicit barrier at the end of the single construct unless a nowait clause is specified. 20 Syntax 21 The syntax of the single construct is as follows: C/C++ #pragma omp single [clause[[,] clause] ...] new-line structured-block 50 OpenMP API • Version 3.1 July 2011 1 where clause is one of the following: private(list) firstprivate(list) copyprivate(list) nowait 2 C/C++ Fortran 3 The syntax of the single construct is as follows: !$omp single [clause[[,] clause] ...] structured-block !$omp end single [end_clause[[,] end_clause] ...] 4 where clause is one of the following: private(list) firstprivate(list) 5 and end_clause is one of the following: copyprivate(list) nowait 6 Fortran 7 8 9 10 11 12 Binding The binding thread set for a single region is the current team. A single region binds to the innermost enclosing parallel region. Only the threads of the team executing the binding parallel region participate in the execution of the structured block and the implied barrier of the single region if the barrier is not eliminated by a nowait clause. Chapter 2 Directives 51 1 Description 2 3 4 The method of choosing a thread to execute the structured block is implementation defined. There is an implicit barrier at the end of the single construct unless a nowait clause is specified. 5 For an example of the single construct, see Section A.14 on page 192. 6 Restrictions 7 Restrictions to the single construct are as follows: 8 • The copyprivate clause must not be used with the nowait clause. 9 • At most one nowait clause can appear on a single construct. C/C++ • A throw executed inside a single region must cause execution to resume within the 10 11 same single region, and the same thread that threw the exception must catch it. C/C++ 12 Cross References 13 • private and firstprivate clauses, see Section 2.9.3 on page 92. 14 • copyprivate clause, see Section 2.9.4.2 on page 109. Fortran 15 2.5.4 workshare Construct 16 Summary 17 18 19 The workshare construct divides the execution of the enclosed structured block into separate units of work, and causes the threads of the team to share the work such that each unit is executed only once by one thread, in the context of its implicit task. 52 OpenMP API • Version 3.1 July 2011 Fortran (cont.) 1 Syntax 2 The syntax of the workshare construct is as follows: !$omp workshare structured-block !$omp end workshare [nowait] 3 The enclosed structured block must consist of only the following: 4 • array assignments 5 • scalar assignments 6 • FORALL statements 7 • FORALL constructs 8 • WHERE statements 9 • WHERE constructs 10 • atomic constructs 11 • critical constructs 12 • parallel constructs 13 14 Statements contained in any enclosed critical construct are also subject to these restrictions. Statements in any enclosed parallel construct are not restricted. 15 Binding 16 17 18 19 20 The binding thread set for a workshare region is the current team. A workshare region binds to the innermost enclosing parallel region. Only the threads of the team executing the binding parallel region participate in the execution of the units of work and the implied barrier of the workshare region if the barrier is not eliminated by a nowait clause. 21 Description 22 23 There is an implicit barrier at the end of a workshare construct unless a nowait clause is specified. Chapter 2 Directives 53 Fortran (cont.) 1 2 3 4 5 An implementation of the workshare construct must insert any synchronization that is required to maintain standard Fortran semantics. For example, the effects of one statement within the structured block must appear to occur before the execution of succeeding statements, and the evaluation of the right hand side of an assignment must appear to complete prior to the effects of assigning to the left hand side. 6 The statements in the workshare construct are divided into units of work as follows: 7 8 • For array expressions within each statement, including transformational array intrinsic functions that compute scalar values from arrays: 9 10 • Evaluation of each element of the array expression, including any references to ELEMENTAL functions, is a unit of work. 11 12 • Evaluation of transformational array intrinsic functions may be freely subdivided into any number of units of work. 13 • For an array assignment statement, the assignment of each element is a unit of work. 14 • For a scalar assignment statement, the assignment operation is a unit of work. 15 16 • For a WHERE statement or construct, the evaluation of the mask expression and the 17 18 19 • For a FORALL statement or construct, the evaluation of the mask expression, 20 21 • For an atomic construct, the atomic operation on the storage location designated as 22 • For a critical construct, the construct is a single unit of work. 23 24 25 • For a parallel construct, the construct is a unit of work with respect to the 26 27 • If none of the rules above apply to a portion of a statement in the structured block, 28 29 30 The transformational array intrinsic functions are MATMUL, DOT_PRODUCT, SUM, PRODUCT, MAXVAL, MINVAL, COUNT, ANY, ALL, SPREAD, PACK, UNPACK, RESHAPE, TRANSPOSE, EOSHIFT, CSHIFT, MINLOC, and MAXLOC. 31 32 It is unspecified how the units of work are assigned to the threads executing a workshare region. 33 34 35 If an array expression in the block references the value, association status, or allocation status of private variables, the value of the expression is undefined, unless the same value would be computed by every thread. masked assignments are each a unit of work. expressions occurring in the specification of the iteration space, and the masked assignments are each a unit of work. x is the unit of work. workshare construct. The statements contained in the parallel construct are executed by a new thread team. then that portion is a unit of work. 54 OpenMP API • Version 3.1 July 2011 1 2 If an array assignment, a scalar assignment, a masked array assignment, or a FORALL assignment assigns to a private variable in the block, the result is unspecified. 3 4 The workshare directive causes the sharing of work to occur only in the workshare construct, and not in the remainder of the workshare region. 5 For examples of the workshare construct, see Section A.17 on page 213. 6 Restrictions 7 The following restrictions apply to the workshare construct: 8 9 • All array assignments, scalar assignments, and masked array assignments must be intrinsic assignments. • The construct must not contain any user defined function calls unless the function is 10 11 ELEMENTAL. Fortran 13 Combined Parallel Worksharing Constructs 14 15 16 17 Combined parallel worksharing constructs are shortcuts for specifying a worksharing construct nested immediately inside a parallel construct. The semantics of these directives are identical to that of explicitly specifying a parallel construct containing one worksharing construct and no other statements. 18 19 20 21 The combined parallel worksharing constructs allow certain clauses that are permitted both on parallel constructs and on worksharing constructs. If a program would have different behavior depending on whether the clause were applied to the parallel construct or to the worksharing construct, then the program’s behavior is unspecified. 22 The following sections describe the combined parallel worksharing constructs: 23 • The parallel loop construct. 24 • The parallel sections construct. 25 • The parallel workshare construct. 12 2.6 Chapter 2 Directives 55 1 2.6.1 Parallel Loop Construct 2 Summary 3 4 The parallel loop construct is a shortcut for specifying a parallel construct containing one or more associated loops and no other statements. 5 Syntax 6 The syntax of the parallel loop construct is as follows: C/C++ #pragma omp parallel for [clause[[,] clause] ...] new-line for-loop where clause can be any of the clauses accepted by the parallel or for directives, except the nowait clause, with identical meanings and restrictions. 7 8 C/C++ Fortran The syntax of the parallel loop construct is as follows: 9 !$omp parallel do [clause[[,] clause] ...] do-loop [!$omp end parallel do] 10 11 where clause can be any of the clauses accepted by the parallel or do directives, with identical meanings and restrictions. 12 13 14 If an end parallel do directive is not specified, an end parallel do directive is assumed at the end of the do-loop. nowait may not be specified on an end parallel do directive. Fortran 56 OpenMP API • Version 3.1 July 2011 1 Description 2 3 The semantics are identical to explicitly specifying a parallel directive immediately followed by a for directive. C/C++ C/C++ Fortran The semantics are identical to explicitly specifying a parallel directive immediately followed by a do directive, and an end do directive immediately followed by an end parallel directive. 4 5 6 Fortran 7 Restrictions 8 The restrictions for the parallel construct and the loop construct apply. 9 Cross References 10 • parallel construct, see Section 2.4 on page 33. 11 • loop construct, see Section 2.5.1 on page 39. 12 • Data attribute clauses, see Section 2.9.3 on page 92. 13 2.6.2 parallel sections Construct 14 Summary 15 16 The parallel sections construct is a shortcut for specifying a parallel construct containing one sections construct and no other statements. Chapter 2 Directives 57 1 Syntax 2 The syntax of the parallel sections construct is as follows: C/C++ #pragma omp parallel sections [clause[[,] clause] ...] new-line { [#pragma omp section new-line] structured-block [#pragma omp section new-line structured-block ] ... } where clause can be any of the clauses accepted by the parallel or sections directives, except the nowait clause, with identical meanings and restrictions. 3 4 C/C++ Fortran The syntax of the parallel sections construct is as follows: 5 !$omp parallel sections [clause[[,] clause] ...] [!$omp section] structured-block [!$omp section structured-block ] ... !$omp end parallel sections 6 7 where clause can be any of the clauses accepted by the parallel or sections directives, with identical meanings and restrictions. 8 9 The last section ends at the end parallel sections directive. nowait cannot be specified on an end parallel sections directive. Fortran 10 Description 11 12 The semantics are identical to explicitly specifying a parallel directive immediately followed by a sections directive. C/C++ C/C++ 58 OpenMP API • Version 3.1 July 2011 Fortran The semantics are identical to explicitly specifying a parallel directive immediately followed by a sections directive, and an end sections directive immediately followed by an end parallel directive. 1 2 3 Fortran 4 For an example of the parallel sections construct, see Section A.12 on page 189. 5 Restrictions 6 The restrictions for the parallel construct and the sections construct apply. 7 Cross References: 8 • parallel construct, see Section 2.4 on page 33. 9 • sections construct, see Section 2.5.2 on page 48. 10 • Data attribute clauses, see Section 2.9.3 on page 92. Fortran 11 2.6.3 parallel workshare Construct 12 Summary 13 14 The parallel workshare construct is a shortcut for specifying a parallel construct containing one workshare construct and no other statements. 15 Syntax 16 The syntax of the parallel workshare construct is as follows: !$omp parallel workshare [clause[[,] clause] ...] structured-block !$omp end parallel workshare 17 18 19 where clause can be any of the clauses accepted by the parallel directive, with identical meanings and restrictions. nowait may not be specified on an end parallel workshare directive. Chapter 2 Directives 59 1 Description 2 3 4 The semantics are identical to explicitly specifying a parallel directive immediately followed by a workshare directive, and an end workshare directive immediately followed by an end parallel directive. 5 Restrictions 6 The restrictions for the parallel construct and the workshare construct apply. 7 Cross References 8 • parallel construct, see Section 2.4 on page 33. 9 • workshare construct, see Section 2.5.4 on page 52. • Data attribute clauses, see Section 2.9.3 on page 92. 10 Fortran 60 OpenMP API • Version 3.1 July 2011 1 2.7 Tasking Constructs 2 2.7.1 task Construct 3 Summary 4 The task construct defines an explicit task. 5 Syntax 6 The syntax of the task construct is as follows: C/C++ #pragma omp task [clause[[,] clause] ...] new-line structured-block 7 where clause is one of the following: if(scalar-expression) final(scalar-expression) untied default(shared | none) mergeable private(list) firstprivate(list) shared(list) C/C++ 8 Chapter 2 Directives 61 Fortran The syntax of the task construct is as follows: 1 !$omp task [clause[[,] clause] ...] structured-block !$omp end task where clause is one of the following: 2 if(scalar-logical-expression) final(scalar-logical-expression) untied default(private | firstprivate | shared | none) mergeable private(list) firstprivate(list) shared(list) Fortran 3 4 Binding 5 6 The binding thread set of the task region is the current team. A task region binds to the innermost enclosing parallel region. 7 Description 8 9 10 11 When a thread encounters a task construct, a task is generated from the code for the associated structured block. The data environment of the task is created according to the data-sharing attribute clauses on the task construct, per-data environment ICVs, and any defaults that apply. 12 13 14 15 16 The encountering thread may immediately execute the task, or defer its execution. In the latter case, any thread in the team may be assigned the task. Completion of the task can be guaranteed using task synchronization constructs. A task construct may be nested inside an outer task, but the task region of the inner task is not a part of the task region of the outer task. 62 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 When an if clause is present on a task construct, and the if clause expression evaluates to false, an undeferred task is generated, and the encountering thread must suspend the current task region, for which execution cannot be resumed until the generated task is completed. Note that the use of a variable in an if clause expression of a task construct causes an implicit reference to the variable in all enclosing constructs. 7 8 9 10 11 When a final clause is present on a task construct and the final clause expression evaluates to true, the generated task will be a final task. All task constructs encountered during execution of a final task will generate final and included tasks. Note that the use of a variable in a final clause expression of a task construct causes an implicit reference to the variable in all enclosing constructs. 12 13 The if clause expression and the final clause expression are evaluated in the context outside of the task construct, and no ordering of those evaluations is specified. 14 15 16 17 18 19 20 A thread that encounters a task scheduling point within the task region may temporarily suspend the task region. By default, a task is tied and its suspended task region can only be resumed by the thread that started its execution. If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. The untied clause is ignored if a final clause is present on the same task construct and the final clause expression evaluates to true, or if a task is an included task. 21 22 23 24 The task construct includes a task scheduling point in the task region of its generating task, immediately following the generation of the explicit task. Each explicit task region includes a task scheduling point at its point of completion. An implementation might add task scheduling points anywhere in untied task regions. 25 26 27 When a mergeable clause is present on a task construct, and the generated task is an undeferred task or an included task, the implementation might generate a merged task instead. 28 29 30 Note – When storage is shared by an explicit task region, it is the programmer's responsibility to ensure, by adding proper synchronization, that the storage does not reach the end of its lifetime before the explicit task region completes its execution. 31 Restrictions 32 Restrictions to the task construct are as follows: 33 • A program that branches into or out of a task region is non-conforming. 34 35 • A program must not depend on any ordering of the evaluations of the clauses of the task directive, or on any side effects of the evaluations of the clauses. Chapter 2 Directives 63 1 • At most one if clause can appear on the directive. 2 • At most one final clause can appear on the directive. 3 4 • A throw executed inside a task region must cause execution to resume within the C/C++ same task region, and the same thread that threw the exception must catch it. C/C++ Fortran • Unsynchronized use of Fortran I/O statements by multiple tasks on the same unit has 5 6 unspecified behavior. Fortran 7 2.7.2 taskyield Construct Summary 8 9 10 The taskyield construct specifies that the current task can be suspended in favor of execution of a different task. 11 Syntax 12 The syntax of the taskyield construct is as follows: C/C++ #pragma omp taskyield new-line Because the taskyield construct is a stand-alone directive, there are some restrictions on its placement within a program. The taskyield directive may be placed only at a point where a base language statement is allowed. The taskyield directive may not be used in place of the statement following an if, while, do, switch, or label. See Appendix C for the formal grammar. The examples in Section A.25 on page 236 illustrate these restrictions. 13 14 15 16 17 18 C/C++ Fortran The syntax of the taskyield construct is as follows: 19 !$omp taskyield 64 OpenMP API • Version 3.1 July 2011 Because the taskyield construct is a stand-alone directive, there are some restrictions on its placement within a program. The taskyield directive may be placed only at a point where a Fortran executable statement is allowed. The taskyield directive may not be used as the action statement in an if statement or as the executable statement following a label if the label is referenced in the program. The examples in Section A.25 on page 236 illustrate these restrictions. 1 2 3 4 5 6 Fortran 7 Binding 8 9 A taskyield region binds to the current task region. The binding thread set of the taskyield region is the current team. 10 Description 11 12 The taskyield region includes an explicit task scheduling point in the current task region. 13 Cross References 14 • Task scheduling, see Section 2.7.3 on page 65. 15 2.7.3 Task Scheduling 16 17 18 Whenever a thread reaches a task scheduling point, the implementation may cause it to perform a task switch, beginning or resuming execution of a different task bound to the current team. Task scheduling points are implied at the following locations: 19 • the point immediately following the generation of an explicit task 20 • after the last instruction of a task region 21 • in taskyield regions 22 • in taskwait regions 23 • in implicit and explicit barrier regions. 24 25 In addition, implementations may insert implementation defined task scheduling points in untied tasks anywhere that they are not specifically prohibited in this specification. 26 27 When a thread encounters a task scheduling point it may do one of the following, subject to the Task Scheduling Constraints (below): 28 • begin execution of a tied task bound to the current team Chapter 2 Directives 65 1 • resume any suspended task region, bound to the current team, to which it is tied 2 • begin execution of an untied task bound to the current team 3 • resume any suspended untied task region bound to the current team. 4 5 If more than one of the above choices is available, it is unspecified as to which will be chosen. 6 Task Scheduling Constraints are as follows: 7 1. An included task is executed immediately after generation of the task. 8 9 10 11 2. Scheduling of new tied tasks is constrained by the set of task regions that are currently tied to the thread, and that are not suspended in a barrier region. If this set is empty, any new tied task may be scheduled. Otherwise, a new tied task may be scheduled only if it is a descendant of every task in the set. 12 13 14 3. When an explicit task is generated by a construct containing an if clause for which the expression evaluated to false, and the previous constraint is already met, the task is executed immediately after generation of the task. 15 A program relying on any other assumption about task scheduling is non-conforming. 16 17 18 19 20 Note – Task scheduling points dynamically divide task regions into parts. Each part is executed uninterrupted from start to end. Different parts of the same task region are executed in the order in which they are encountered. In the absence of task synchronization constructs, the order in which a thread executes parts of different schedulable tasks is unspecified. 21 22 A correct program must behave correctly and consistently with all conceivable scheduling sequences that are compatible with the rules above. 23 24 25 26 27 For example, if threadprivate storage is accessed (explicitly in the source code or implicitly in calls to library routines) in one part of a task region, its value cannot be assumed to be preserved into the next part of the same task region if another schedulable task exists that modifies it (see Example A.15.7c on page 202, Example A.15.7f on page 202, Example A.15.8c on page 203 and Example A.15.8f on page 203). 28 29 30 31 32 33 34 As another example, if a lock acquire and release happen in different parts of a task region, no attempt should be made to acquire the same lock in any part of another task that the executing thread may schedule. Otherwise, a deadlock is possible. A similar situation can occur when a critical region spans multiple parts of a task and another schedulable task contains a critical region with the same name (see Example A.15.9c on page 204, Example A.15.9f on page 205, Example A.15.10c on page 206 and Example A.15.10f on page 207). 66 OpenMP API • Version 3.1 July 2011 The use of threadprivate variables and the use of locks or critical sections in an explicit task with an if clause must take into account that when the if clause evaluates to false, the task is executed immediately, without regard to Task Scheduling Constraint 2. 1 2 3 4 2.8 Master and Synchronization Constructs 5 The following sections describe : 6 • the master construct. 7 • the critical construct. 8 • the barrier construct. 9 • the taskwait construct. 10 • the atomic construct. 11 • the flush construct. 12 • the ordered construct. 13 2.8.1 master Construct 14 Summary 15 16 The master construct specifies a structured block that is executed by the master thread of the team. 17 Syntax 18 The syntax of the master construct is as follows: C/C++ #pragma omp master new-line structured-block 19 C/C++ Chapter 2 Directives 67 Fortran The syntax of the master construct is as follows: 1 !$omp master structured-block !$omp end master 2 Fortran 3 Binding 4 5 6 7 The binding thread set for a master region is the current team. A master region binds to the innermost enclosing parallel region. Only the master thread of the team executing the binding parallel region participates in the execution of the structured block of the master region. 8 Description 9 10 Other threads in the team do not execute the associated structured block. There is no implied barrier either on entry to, or exit from, the master construct. 11 For an example of the master construct, see Section A.18 on page 217. 12 Restrictions 13 14 • A throw executed inside a master region must cause execution to resume within the C/C++ same master region, and the same thread that threw the exception must catch it. C/C++ 15 2.8.2 critical Construct 16 Summary 17 18 The critical construct restricts execution of the associated structured block to a single thread at a time. 68 OpenMP API • Version 3.1 July 2011 1 Syntax 2 The syntax of the critical construct is as follows: C/C++ #pragma omp critical [(name)] new-line structured-block 3 C/C++ Fortran 4 The syntax of the critical construct is as follows: !$omp critical [(name)] structured-block !$omp end critical [(name)] 5 Fortran 6 Binding 7 8 9 The binding thread set for a critical region is all threads. Region execution is restricted to a single thread at a time among all the threads in the program, without regard to the team(s) to which the threads belong. 10 Description 11 12 13 14 15 16 An optional name may be used to identify the critical construct. All critical constructs without a name are considered to have the same unspecified name. A thread waits at the beginning of a critical region until no thread is executing a critical region with the same name. The critical construct enforces exclusive access with respect to all critical constructs with the same name in all threads, not just those threads in the current team. 17 18 19 Identifiers used to identify a critical construct have external linkage and are in a name space that is separate from the name spaces used by labels, tags, members, and ordinary identifiers. C/C++ C/C++ Chapter 2 Directives 69 Fortran The names of critical constructs are global entities of the program. If a name conflicts with any other entity, the behavior of the program is unspecified. 1 2 Fortran 3 For an example of the critical construct, see Section A.19 on page 219. 4 Restrictions 5 6 7 • A throw executed inside a critical region must cause execution to resume within C/C++ the same critical region, and the same thread that threw the exception must catch it. C/C++ Fortran The following restrictions apply to the critical construct: 8 9 10 • If a name is specified on a critical directive, the same name must also be 11 12 • If no name appears on the critical directive, no name can appear on the end specified on the end critical directive. critical directive. Fortran 13 2.8.3 barrier Construct 14 Summary 15 16 The barrier construct specifies an explicit barrier at the point at which the construct appears. 17 Syntax 18 The syntax of the barrier construct is as follows: C/C++ #pragma omp barrier new-line 70 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 Because the barrier construct is a stand-alone directive, there are some restrictions on its placement within a program. The barrier directive may be placed only at a point where a base language statement is allowed. The barrier directive may not be used in place of the statement following an if, while, do, switch, or label. See Appendix C for the formal grammar. The examples in Section A.25 on page 236 illustrate these restrictions. C/C++ Fortran 7 The syntax of the barrier construct is as follows: !$omp barrier 8 9 10 11 12 13 14 Because the barrier construct is a stand-alone directive, there are some restrictions on its placement within a program. The barrier directive may be placed only at a point where a Fortran executable statement is allowed. The barrier directive may not be used as the action statement in an if statement or as the executable statement following a label if the label is referenced in the program. The examples in Section A.25 on page 236 illustrate these restrictions. Fortran 15 Binding 16 17 18 The binding thread set for a barrier region is the current team. A barrier region binds to the innermost enclosing parallel region. See Section A.21 on page 222 for examples. 19 Description 20 21 22 23 All threads of the team executing the binding parallel region must execute the barrier region and complete execution of all explicit tasks generated in the binding parallel region up to this point before any are allowed to continue execution beyond the barrier. 24 25 The barrier region includes an implicit task scheduling point in the current task region. 26 Restrictions 27 The following restrictions apply to the barrier construct: Chapter 2 Directives 71 1 • Each barrier region must be encountered by all threads in a team or by none at all. 2 3 • The sequence of worksharing regions and barrier regions encountered must be the 4 same for every thread in a team. 2.8.4 taskwait Construct 5 Summary 6 7 The taskwait construct specifies a wait on the completion of child tasks of the current task. 8 Syntax 9 The syntax of the taskwait construct is as follows: C/C++ #pragma omp taskwait newline Because the taskwait construct is a stand-alone directive, there are some restrictions on its placement within a program. The taskwait directive may be placed only at a point where a base language statement is allowed. The taskwait directive may not be used in place of the statement following an if, while, do, switch, or label. See Appendix C for the formal grammar. The examples in Section A.25 on page 236 illustrate these restrictions. 10 11 12 13 14 15 C/C++ Fortran The syntax of the taskwait construct is as follows: 16 !$omp taskwait Because the taskwait construct is a stand-alone directive, there are some restrictions on its placement within a program. The taskwait directive may be placed only at a point where a Fortran executable statement is allowed. The taskwait directive may not be used as the action statement in an if statement or as the executable statement following a label if the label is referenced in the program. The examples in Section A.25 on page 236 illustrate these restrictions. 17 18 19 20 21 22 Fortran 72 OpenMP API • Version 3.1 July 2011 1 Binding 2 3 A taskwait region binds to the current task region. The binding thread set of the taskwait region is the current team. 4 Description 5 6 7 The taskwait region includes an implicit task scheduling point in the current task region. The current task region is suspended at the task scheduling point until execution of all its child tasks generated before the taskwait region are completed. 8 9 2.8.5 atomic Construct Summary 10 11 12 The atomic construct ensures that a specific storage location is accessed atomically, rather than exposing it to the possibility of multiple, simultaneous reading and writing threads that may result in indeterminate values. 13 Syntax 14 The syntax of the atomic construct takes either of the following forms: C/C++ #pragma omp atomic [read | write | update | capture ] new-line expression-stmt 15 or: #pragma omp atomic capture new-line structured-block 16 where expression-stmt is an expression statement with one of the following forms: 17 18 • If clause is read: 19 20 • If clause is write: v = x; x = expr; Chapter 2 Directives 73 • If clause is update or not present: 1 2 3 4 5 6 7 x++; x--; ++x; --x; x binop= expr; x = x binop expr; 8 9 10 11 12 13 • If clause is capture: 14 and where structured-block is a structured block with one of the following forms: v v v v v = = = = = x++; x--; ++x; --x; x binop= expr; {v = x; x binop= expr;} {x binop= expr; v = x;} {v = x; x = x binop expr;} {x = x binop expr; v = x;} {v = x; x++;} {v = x; ++x;} {++x; v = x;} {x++; v = x;} {v = x; x--;} {v = x; --x;} {--x; v = x;} {x--; v = x;} 15 16 17 18 19 20 21 22 23 24 25 26 27 In the preceding expressions: 28 • x and v (as applicable) are both l-value expressions with scalar type. 29 30 • During the execution of an atomic region, multiple syntactic occurrences of x must 31 • Neither of v and expr (as applicable) may access the storage location designated by x. 32 • Neither of x and expr (as applicable) may access the storage location designated by v. 33 • expr is an expression with scalar type. 34 • binop is one of +, *, -, /, &, ^, |, <<, or >>. 35 • binop, binop=, ++, and -- are not overloaded operators. 36 37 • For forms that allow multiple occurrences of x, the number of times that x is designate the same storage location. evaluated is unspecified. C/C++ 74 OpenMP API • Version 3.1 July 2011 Fortran 1 The syntax of the atomic construct takes any of the following forms: !$omp atomic read capture-statement [!$omp end atomic] 2 or !$omp atomic write write-statement [!$omp end atomic] 3 or !$omp atomic [update] update-statement [!$omp end atomic] 4 or !$omp atomic capture update-statement capture-statement !$omp end atomic 5 or !$omp atomic capture capture-statement update-statement !$omp end atomic 6 7 8 9 10 11 where write-statement has the following form (if clause is write): x = expr where capture-statement has the following form (if clause is capture or read): v=x and where update-statement has one of the following forms (if clause is update, capture, or not present): Chapter 2 Directives 75 1 x = x operator expr 2 x = expr operator x 3 x = intrinsic_procedure_name (x, expr_list) 4 x = intrinsic_procedure_name (expr_list, x) 5 In the preceding statements: 6 • x and v (as applicable) are both scalar variables of intrinsic type. 7 8 • During the execution of an atomic region, multiple syntactic occurrences of x must designate the same storage location. 9 10 • None of v, expr and expr_list (as applicable) may access the same storage location as 11 12 • None of x, expr and expr_list (as applicable) may access the same storage location as 13 • expr is a scalar expression. 14 15 16 • expr_list is a comma-separated, non-empty list of scalar expressions. If 17 • intrinsic_procedure_name is one of MAX, MIN, IAND, IOR, or IEOR. 18 • operator is one of +, *, -, /, .AND., .OR., .EQV., or .NEQV. . 19 20 21 • The operators in expr must have precedence equal to or greater than the precedence 22 23 • intrinsic_procedure_name must refer to the intrinsic procedure name and not to other 24 • operator must refer to the intrinsic operator and not to a user-defined operator. 25 • All assignments must be intrinsic assignments. 26 27 • For forms that allow multiple occurrences of x, the number of times that x is x. v. intrinsic_procedure_name refers to IAND, IOR, or IEOR, exactly one expression must appear in expr_list. of operator, x operator expr must be mathematically equivalent to x operator (expr), and expr operator x must be mathematically equivalent to (expr) operator x. program entities. evaluated is unspecified. Fortran 28 Binding 29 30 31 32 The binding thread set for an atomic region is all threads. atomic regions enforce exclusive access with respect to other atomic regions that access the same storage location x among all the threads in the program without regard to the teams to which the threads belong. 76 OpenMP API • Version 3.1 July 2011 1 Description 2 3 The atomic construct with the read clause forces an atomic read of the location designated by x regardless of the native machine word size. 4 5 The atomic construct with the write clause forces an atomic write of the location designated by x regardless of the native machine word size. 6 7 8 9 10 11 12 The atomic construct with the update clause forces an atomic update of the location designated by x using the designated operator or intrinsic. Note that when no clause is present, the semantics are equivalent to atomic update. Only the read and write of the location designated by x are performed mutually atomically. The evaluation of expr or expr_list need not be atomic with respect to the read or write of the location designated by x. No task scheduling points are allowed between the read and the write of the location designated by x. 13 14 15 16 17 18 19 20 21 22 The atomic construct with the capture clause forces an atomic update of the location designated by x using the designated operator or intrinsic while also capturing the original or final value of the location designated by x with respect to the atomic update. The original or final value of the location designated by x is written in the location designated by v depending on the form of the atomic construct structured block or statements following the usual language semantics. Only the read and write of the location designated by x are performed mutually atomically. Neither the evaluation of expr or expr_list, nor the write to the location designated by v need be atomic with respect to the read or write of the location designated by x. No task scheduling points are allowed between the read and the write of the location designated by x. 23 24 25 26 For all forms of the atomic construct, any combination of two or more of these atomic constructs enforces mutually exclusive access to the locations designated by x. To avoid race conditions, all accesses of the locations designated by x that could potentially occur in parallel must be protected with an atomic construct. 27 28 29 30 atomic regions do not guarantee exclusive access with respect to any accesses outside of atomic regions to the same storage location x even if those accesses occur during a critical or ordered region, while an OpenMP lock is owned by the executing task, or during the execution of a reduction clause. 31 32 33 However, other OpenMP synchronization can ensure the desired exclusive access. For example, a barrier following a series of atomic updates to x guarantees that subsequent accesses do not form a race with the atomic accesses. 34 35 36 A compliant implementation may enforce exclusive access between atomic regions that update different storage locations. The circumstances under which this occurs are implementation defined. 37 For an example of the atomic construct, see Section A.22 on page 224. 38 Chapter 2 Directives 77 1 Restrictions 2 The following restriction applies to the atomic construct: 3 4 • All atomic accesses to the storage locations designated by x throughout the program C/C++ are required to have a compatible type. See Section A.23 on page 230 for examples. C/C++ Fortran 5 The following restriction applies to the atomic construct: 6 7 8 • All atomic accesses to the storage locations designated by x throughout the program are required to have the same type and type parameters. See Section A.23 on page 230 for examples. Fortran Cross References 9 10 • critical construct, see Section 2.8.2 on page 68. 11 • barrier construct, see Section 2.8.3 on page 70. 12 • flush construct, see Section 2.8.6 on page 78. 13 • ordered construct, see Section 2.8.7 on page 82. 14 • reduction clause, see Section 2.9.3.6 on page 103. 15 • lock routines, see Section 3.3 on page 141. 16 2.8.6 flush Construct 17 Summary 18 19 20 21 The flush construct executes the OpenMP flush operation. This operation makes a thread’s temporary view of memory consistent with memory, and enforces an order on the memory operations of the variables explicitly specified or implied. See the memory model description in Section 1.4 on page 13 for more details. 78 OpenMP API • Version 3.1 July 2011 1 Syntax 2 The syntax of the flush construct is as follows: C/C++ #pragma omp flush [(list)] new-line 3 4 5 6 7 8 Because the flush construct is a stand-alone directive, there are some restrictions on its placement within a program. The flush directive may be placed only at a point where a base language statement is allowed. The flush directive may not be used in place of the statement following an if, while, do, switch, or label. See Appendix C for the formal grammar. See Section A.25 on page 236 for an example that illustrates these placement restrictions. C/C++ Fortran 9 The syntax of the flush construct is as follows: !$omp flush [(list)] 10 11 12 13 14 15 Because the flush construct is a stand-alone directive, there are some restrictions on its placement within a program. The flush directive may be placed only at a point where a Fortran executable statement is allowed. The flush directive may not be used as the action statement in an if statement or as the executable statement following a label if the label is referenced in the program. The examples in Section A.25 on page 236 illustrate these restrictions. Fortran 16 Binding 17 18 19 20 21 The binding thread set for a flush region is the encountering thread. Execution of a flush region affects the memory and the temporary view of memory of only the thread that executes the region. It does not affect the temporary view of other threads. Other threads must themselves execute a flush operation in order to be guaranteed to observe the effects of the encountering thread’s flush operation. 22 Description 23 24 25 A flush construct without a list, executed on a given thread, operates as if the whole thread-visible data state of the program, as defined by the base language, is flushed. A flush construct with a list applies the flush operation to the items in the list, and does Chapter 2 Directives 79 1 2 3 4 not return until the operation is complete for all specified list items. Use of a flush construct with a list is extremely error prone and users are strongly discouraged from attempting it. An implementation may implement a flush with a list by ignoring the list, and treating it the same as a flush without a list. 5 6 If a pointer is present in the list, the pointer itself is flushed, not the memory block to which the pointer refers. C/C++ C/C++ Fortran If the list item or a subobject of the list item has the POINTER attribute, the allocation or association status of the POINTER item is flushed, but the pointer target is not. If the list item is a Cray pointer, the pointer is flushed, but the object to which it points is not. If the list item has the ALLOCATABLE attribute and the list item is allocated, the allocated array is flushed; otherwise the allocation status is flushed. 7 8 9 10 11 Fortran 12 For examples of the flush construct, see Section A.25 on page 236. 13 14 15 16 17 18 19 20 21 Note – the following examples illustrate the ordering properties of the flush operation. In the following incorrect pseudocode example, the programmer intends to prevent simultaneous execution of the protected section by the two threads, but the program does not work properly because it does not enforce the proper ordering of the operations on variables a and b. Any shared data accessed in the protected section is not guaranteed to be current or consistent during or after the protected section. The atomic notation in the pseudocode in the following two examples indicates that the accesses to a and b are ATOMIC writes and captures. Otherwise both examples would contain data races and automatically result in unspecified behavior. Incorrect example: a = b = 0 thread 1 atomic(b = 1) flush(b) flush(a) atomic(tmp = a) if (tmp == 0) then protected section end if 80 OpenMP API • Version 3.1 July 2011 thread 2 atomic(a = 1) flush(a) flush(b) atomic(tmp = b) if (tmp == 0) then protected section end if 1 2 3 4 5 6 The problem with this example is that operations on variables a and b are not ordered with respect to each other. For instance, nothing prevents the compiler from moving the flush of b on thread 1 or the flush of a on thread 2 to a position completely after the protected section (assuming that the protected section on thread 1 does not reference b and the protected section on thread 2 does not reference a). If either re-ordering happens, both threads can simultaneously execute the protected section. 7 8 9 10 The following pseudocode example correctly ensures that the protected section is executed by not more than one of the two threads at any one time. Notice that execution of the protected section by neither thread is considered correct in this example. This occurs if both flushes complete prior to either thread executing its if statement. Correct example: a = b = 0 thread 1 atomic(b = 1) flush(a,b) atomic(tmp = a) if (tmp == 0) then protected section end if thread 2 atomic(a = 1) flush(a,b) atomic(tmp = b) if (tmp == 0) then protected section end if 11 12 13 The compiler is prohibited from moving the flush at all for either thread, ensuring that the respective assignment is complete and the data is flushed before the if statement is executed. 14 A flush region without a list is implied at the following locations: 15 • During a barrier region. 16 • At entry to and exit from parallel, critical, and ordered regions. 17 • At exit from worksharing regions unless a nowait is present. 18 • At entry to and exit from combined parallel worksharing regions. 19 • During omp_set_lock and omp_unset_lock regions. 20 21 22 • During omp_test_lock, omp_set_nest_lock, omp_unset_nest_lock 23 • Immediately before and immediately after every task scheduling point. 24 A flush region with a list is implied at the following locations: and omp_test_nest_lock regions, if the region causes the lock to be set or unset. Chapter 2 Directives 81 1 2 3 4 5 • At entry to and exit from the atomic operation (read, write, update, or capture) 6 Note – A flush region is not implied at the following locations: 7 • At entry to worksharing regions. 8 • At entry to or exit from a master region. 9 performed in an atomic region, where the list contains only the storage location designated as x according to the description of the syntax of the atomic construct in Section 2.8.5 on page 73. 2.8.7 ordered Construct 10 Summary 11 12 13 The ordered construct specifies a structured block in a loop region that will be executed in the order of the loop iterations. This sequentializes and orders the code within an ordered region while allowing code outside the region to run in parallel. 14 Syntax 15 The syntax of the ordered construct is as follows: C/C++ #pragma omp ordered new-line structured-block C/C++ 16 Fortran The syntax of the ordered construct is as follows: 17 !$omp ordered structured-block !$omp end ordered 18 Fortran 82 OpenMP API • Version 3.1 July 2011 1 Binding 2 3 4 The binding thread set for an ordered region is the current team. An ordered region binds to the innermost enclosing loop region. ordered regions that bind to different loop regions execute independently of each other. 5 Description 6 7 8 9 10 11 The threads in the team executing the loop region execute ordered regions sequentially in the order of the loop iterations. When the thread executing the first iteration of the loop encounters an ordered construct, it can enter the ordered region without waiting. When a thread executing any subsequent iteration encounters an ordered region, it waits at the beginning of that ordered region until execution of all the ordered regions belonging to all previous iterations have completed. 12 For examples of the ordered construct, see Section A.26 on page 239. 13 Restrictions 14 Restrictions to the ordered construct are as follows: 15 16 • The loop region to which an ordered region binds must have an ordered clause 17 18 19 • During execution of an iteration of a loop or a loop nest within a loop region, a 20 21 22 • A throw executed inside a ordered region must cause execution to resume within specified on the corresponding loop (or parallel loop) construct. thread must not execute more than one ordered region that binds to the same loop region. C/C++ the same ordered region, and the same thread that threw the exception must catch it. C/C++ 23 Cross References 24 • loop construct, see Section 2.5.1 on page 39. 25 • parallel loop construct, see Section 2.6.1 on page 56. Chapter 2 Directives 83 1 2.9 Data Environment 2 3 This section presents a directive and several clauses for controlling the data environment during the execution of parallel, task, and worksharing regions. 4 5 • Section 2.9.1 on page 84 describes how the data-sharing attributes of variables 6 7 • The threadprivate directive, which is provided to create threadprivate memory, referenced in parallel, task, and worksharing regions are determined. is described in Section 2.9.2 on page 88. 8 9 10 • Clauses that may be specified on directives to control the data-sharing attributes of 11 12 13 • Clauses that may be specified on directives to copy data values from private or 14 variables referenced in parallel, task, or worksharing constructs are described in Section 2.9.3 on page 92. threadprivate variables on one thread to the corresponding variables on other threads in the team are described in Section 2.9.4 on page 107. 2.9.1 Data-sharing Attribute Rules 15 16 17 This section describes how the data-sharing attributes of variables referenced in parallel, task, and worksharing regions are determined. The following two cases are described separately: 18 19 • Section 2.9.1.1 on page 84 describes the data-sharing attribute rules for variables 20 21 • Section 2.9.1.2 on page 87 describes the data-sharing attribute rules for variables 22 23 referenced in a construct. referenced in a region, but outside any construct. 2.9.1.1 Data-sharing Attribute Rules for Variables Referenced in a Construct 24 25 26 The data-sharing attributes of variables that are referenced in a construct can be predetermined, explicitly determined, or implicitly determined, according to the rules outlined in this section. 27 28 29 30 Specifying a variable on a firstprivate, lastprivate, or reduction clause of an enclosed construct causes an implicit reference to the variable in the enclosing construct. Such implicit references are also subject to the data-sharing attribute rules outlined in this section. 31 Certain variables and objects have predetermined data-sharing attributes as follows: 84 OpenMP API • Version 3.1 July 2011 C/C++ 1 • Variables appearing in threadprivate directives are threadprivate. 2 3 • Variables with automatic storage duration that are declared in a scope inside the 4 • Objects with dynamic storage duration are shared. 5 • Static data members are shared. 6 7 • The loop iteration variable(s) in the associated for-loop(s) of a for or parallel 8 • Variables with const-qualified type having no mutable member are shared. 9 10 construct are private. for construct is (are) private. • Variables with static storage duration that are declared in a scope inside the construct are shared. C/C++ Fortran 11 12 • Variables and common blocks appearing in threadprivate directives are 13 14 • The loop iteration variable(s) in the associated do-loop(s) of a do or parallel do 15 16 • A loop iteration variable for a sequential loop in a parallel or task construct is 17 • Implied-do indices and forall indices are private. 18 19 • Cray pointees inherit the data-sharing attribute of the storage with which their Cray 20 • Assumed-size arrays are shared. threadprivate. construct is(are) private. private in the innermost such construct that encloses the loop. pointers are associated. Fortran 21 22 23 24 Variables with predetermined data-sharing attributes may not be listed in data-sharing attribute clauses, except for the cases listed below. For these exceptions only, listing a predetermined variable in a data-sharing attribute clause is allowed and overrides the variable’s predetermined data-sharing attributes. 25 26 • The loop iteration variable(s) in the associated for-loop(s) of a for or parallel 27 28 • Variables with const-qualified type having no mutable member may be listed in a C/C++ for construct may be listed in a private or lastprivate clause. firstprivate clause. C/C++ Chapter 2 Directives 85 Fortran 1 2 • The loop iteration variable(s) in the associated do-loop(s) of a do or parallel do 3 4 5 • Variables used as loop iteration variables in sequential loops in a parallel or 6 • Assumed-size arrays may be listed in a shared clause. construct may be listed in a private or lastprivate clause. task construct may be listed in data-sharing clauses on the construct itself, and on enclosed constructs, subject to other restrictions. Fortran Additional restrictions on the variables that may appear in individual clauses are described with each clause in Section 2.9.3 on page 92. 7 8 9 10 Variables with explicitly determined data-sharing attributes are those that are referenced in a given construct and are listed in a data-sharing attribute clause on the construct. 11 12 13 Variables with implicitly determined data-sharing attributes are those that are referenced in a given construct, do not have predetermined data-sharing attributes, and are not listed in a data-sharing attribute clause on the construct. 14 Rules for variables with implicitly determined data-sharing attributes are as follows: 15 16 • In a parallel or task construct, the data-sharing attributes of these variables are 17 18 • In a parallel construct, if no default clause is present, these variables are 19 20 • For constructs other than task, if no default clause is present, these variables 21 22 23 • In a task construct, if no default clause is present, a variable that in the determined by the default clause, if present (see Section 2.9.3.1 on page 93). shared. inherit their data-sharing attributes from the enclosing context. enclosing context is determined to be shared by all implicit tasks bound to the current team is shared. Fortran • In an orphaned task construct, if no default clause is present, dummy arguments 24 25 are firstprivate. Fortran 26 27 • In a task construct, if no default clause is present, a variable whose data-sharing 28 29 Additional restrictions on the variables for which data-sharing attributes cannot be implicitly determined in a task construct are described in Section 2.9.3.4 on page 98. attribute is not determined by the rules above is firstprivate. 86 OpenMP API • Version 3.1 July 2011 1 2 2.9.1.2 Data-sharing Attribute Rules for Variables Referenced in a Region but not in a Construct 3 4 The data-sharing attributes of variables that are referenced in a region, but not in a construct, are determined as follows: 5 6 • Variables with static storage duration that are declared in called routines in the region 7 8 • Variables with const-qualified type having no mutable member, and that are C/C++ are shared. declared in called routines, are shared. 9 10 • File-scope or namespace-scope variables referenced in called routines in the region 11 • Objects with dynamic storage duration are shared. 12 • Static data members are shared unless they appear in a threadprivate directive. 13 14 • Formal arguments of called routines in the region that are passed by reference inherit 15 • Other variables declared in called routines in the region are private. are shared unless they appear in a threadprivate directive. the data-sharing attributes of the associated actual argument. C/C++ Fortran 16 17 18 • Local variables declared in called routines in the region and that have the save 19 20 21 • Variables belonging to common blocks, or declared in modules, and referenced in 22 23 • Dummy arguments of called routines in the region that are passed by reference inherit 24 25 • Cray pointees inherit the data-sharing attribute of the storage with which their Cray 26 27 • Implied-do indices, forall indices, and other local variables declared in called attribute, or that are data initialized, are shared unless they appear in a threadprivate directive. called routines in the region are shared unless they appear in a threadprivate directive. the data-sharing attributes of the associated actual argument. pointers are associated. routines in the region are private. Fortran Chapter 2 Directives 87 1 2.9.2 threadprivate Directive 2 Summary 3 4 The threadprivate directive specifies that variables are replicated, with each thread having its own copy. 5 Syntax 6 The syntax of the threadprivate directive is as follows: C/C++ #pragma omp threadprivate(list) new-line where list is a comma-separated list of file-scope, namespace-scope, or static block-scope variables that do not have incomplete types. 7 8 C/C++ Fortran The syntax of the threadprivate directive is as follows: 9 !$omp threadprivate(list) where list is a comma-separated list of named variables and named common blocks. Common block names must appear between slashes. 10 11 Fortran 12 Description 13 14 15 16 17 Each copy of a threadprivate variable is initialized once, in the manner specified by the program, but at an unspecified point in the program prior to the first reference to that copy. The storage of all copies of a threadprivate variable is freed according to how static variables are handled in the base language, but at an unspecified point in the program. 18 19 A program in which a thread references another thread’s copy of a threadprivate variable is non-conforming. 88 OpenMP API • Version 3.1 July 2011 1 2 3 The content of a threadprivate variable can change across a task scheduling point if the executing thread switches to another task that modifies the variable. For more details on task scheduling, see Section 1.3 on page 12 and Section 2.7 on page 61. 4 5 In parallel regions, references by the master thread will be to the copy of the variable in the thread that encountered the parallel region. 6 7 8 During the sequential part references will be to the initial thread’s copy of the variable. The values of data in the initial thread’s copy of a threadprivate variable are guaranteed to persist between any two consecutive references to the variable in the program. 9 10 11 The values of data in the threadprivate variables of non-initial threads are guaranteed to persist between two consecutive active parallel regions only if all the following conditions hold: 12 • Neither parallel region is nested inside another explicit parallel region. 13 • The number of threads used to execute both parallel regions is the same. 14 15 • The value of the dyn-var internal control variable in the enclosing task region is false 16 17 18 If these conditions all hold, and if a threadprivate variable is referenced in both regions, then threads with the same thread number in their respective regions will reference the same copy of that variable. 19 20 21 22 If the above conditions hold, the storage duration, lifetime, and value of a thread’s copy of a threadprivate variable that does not appear in any copyin clause on the second region will be retained. Otherwise, the storage duration, lifetime, and value of a thread’s copy of the variable in the second region is unspecified. 23 24 25 If the value of a variable referenced in an explicit initializer of a threadprivate variable is modified prior to the first reference to any instance of the threadprivate variable, then the behavior is unspecified. 26 27 28 The order in which any constructors for different threadprivate variables of class type are called is unspecified. The order in which any destructors for different threadprivate variables of class type are called is unspecified. at entry to both parallel regions. C/C++ C/C++ 29 Fortran 30 31 A variable is affected by a copyin clause if the variable appears in the copyin clause or it is in a common block that appears in the copyin clause. 32 33 34 If the above conditions hold, the definition, association, or allocation status of a thread’s copy of a threadprivate variable or a variable in a threadprivate common block, that is not affected by any copyin clause that appears on the second region, will Chapter 2 Directives 89 1 2 3 be retained. Otherwise, the definition and association status of a thread’s copy of the variable in the second region is undefined, and the allocation status of an allocatable array will be implementation defined. 4 5 6 7 If a threadprivate variable or a variable in a threadprivate common block is not affected by any copyin clause that appears on the first parallel region in which it is referenced, the variable or any subobject of the variable is initially defined or undefined according to the following rules: 8 9 • If it has the ALLOCATABLE attribute, each copy created will have an initial allocation status of not currently allocated. • If it has the POINTER attribute: 10 11 12 13 • if it has an initial association status of disassociated, either through explicit initialization or default initialization, each copy created will have an association status of disassociated; 14 • otherwise, each copy created will have an association status of undefined. 15 • If it does not have either the POINTER or the ALLOCATABLE attribute: 16 17 • if it is initially defined, either through explicit initialization or default initialization, each copy created is so defined; 18 • otherwise, each copy created is undefined. Fortran 19 For examples of the threadprivate directive, see Section A.27 on page 244. 20 Restrictions 21 The restrictions to the threadprivate directive are as follows: 22 23 • A threadprivate variable must not appear in any clause except the copyin, 24 • A program in which an untied task accesses threadprivate storage is non-conforming. 25 26 27 • A variable that is part of another variable (as an array or structure element) cannot 28 29 30 • A threadprivate directive for file-scope variables must appear outside any 31 32 33 • A threadprivate directive for static class member variables must appear in the copyprivate, schedule, num_threads, and if clauses. C/C++ appear in a threadprivate clause unless it is a static data member of a C++ class. definition or declaration, and must lexically precede all references to any of the variables in its list. class definition, in the same scope in which the member variables are declared, and must lexically precede all references to any of the variables in its list. 90 OpenMP API • Version 3.1 July 2011 1 2 3 • A threadprivate directive for namespace-scope variables must appear outside 4 5 6 • Each variable in the list of a threadprivate directive at file, namespace, or class 7 8 9 • A threadprivate directive for static block-scope variables must appear in the any definition or declaration other than the namespace definition itself, and must lexically precede all references to any of the variables in its list. scope must refer to a variable declaration at file, namespace, or class scope that lexically precedes the directive. scope of the variable and not in a nested scope. The directive must lexically precede all references to any of the variables in its list. 10 11 12 • Each variable in the list of a threadprivate directive in block scope must refer to 13 14 15 • If a variable is specified in a threadprivate directive in one translation unit, it 16 • The address of a threadprivate variable is not an address constant. 17 • A threadprivate variable must not have an incomplete type or a reference type. 18 • A threadprivate variable with class type must have: a variable declaration in the same scope that lexically precedes the directive. The variable declaration must use the static storage-class specifier. must be specified in a threadprivate directive in every translation unit in which it is declared. 19 20 • an accessible, unambiguous default constructor in case of default initialization without a given initializer; 21 22 • an accessible, unambiguous constructor accepting the given argument in case of direct initialization; 23 24 • an accessible, unambiguous copy constructor in case of copy initialization with an explicit initializer. C/C++ 25 Fortran 26 27 • A variable that is part of another variable (as an array or structure element) cannot 28 29 30 31 32 33 • The threadprivate directive must appear in the declaration section of a scoping 34 35 36 37 • If a threadprivate directive specifying a common block name appears in one appear in a threadprivate clause. unit in which the common block or variable is declared. Although variables in common blocks can be accessed by use association or host association, common block names cannot. This means that a common block name specified in a threadprivate directive must be declared to be a common block in the same scoping unit in which the threadprivate directive appears. program unit, then such a directive must also appear in every other program unit that contains a COMMON statement specifying the same name. It must appear after the last such COMMON statement in the program unit. Chapter 2 Directives 91 1 • A blank common block cannot appear in a threadprivate directive. 2 3 4 • A variable can only appear in a threadprivate directive in the scope in which it 5 6 • A variable that appears in a threadprivate directive must be declared in the is declared. It must not be an element of a common block or appear in an EQUIVALENCE statement. scope of a module or have the SAVE attribute, either explicitly or implicitly. Fortran 7 Cross References: 8 • dyn-var ICV, see Section 2.3 on page 28. 9 • number of threads used to execute a parallel region, see Section 2.4.1 on page 36. • copyin clause, see Section 2.9.4.1 on page 107. 10 11 2.9.3 Data-Sharing Attribute Clauses 12 13 14 Several constructs accept clauses that allow a user to control the data-sharing attributes of variables referenced in the construct. Data-sharing attribute clauses apply only to variables for which the names are visible in the construct on which the clause appears. 15 16 Not all of the clauses listed in this section are valid on all directives. The set of clauses that is valid on a particular directive is described with the directive. 17 18 19 20 21 22 Most of the clauses accept a comma-separated list of list items (see Section 2.1 on page 22). All list items appearing in a clause must be visible, according to the scoping rules of the base language. With the exception of the default clause, clauses may be repeated as needed. A list item that specifies a given variable may not appear in more than one clause on the same directive, except that a variable may be specified in both firstprivate and lastprivate clauses. 23 24 25 If a variable referenced in a data-sharing attribute clause has a type derived from a template, and there are no other references to that variable in the program, then any behavior related to that variable is unspecified. C/C++ C/C++ Fortran A named common block may be specified in a list by enclosing the name in slashes. When a named common block appears in a list, it has the same meaning as if every explicit member of the common block appeared in the list. An explicit member of a 26 27 28 92 OpenMP API • Version 3.1 July 2011 1 2 3 common block is a variable that is named in a COMMON statement that specifies the common block name and is declared in the same scoping unit in which the clause appears. 4 5 6 7 Although variables in common blocks can be accessed by use association or host association, common block names cannot. As a result, a common block name specified in a data-sharing attribute clause must be declared to be a common block in the same scoping unit in which the data-sharing attribute clause appears. 8 9 10 11 12 13 14 When a named common block appears in a private, firstprivate, lastprivate, or shared clause of a directive, none of its members may be declared in another data-sharing attribute clause in that directive (see Section A.29 on page 251 for examples). When individual members of a common block appear in a private, firstprivate, lastprivate, or reduction clause of a directive, the storage of the specified variables is no longer associated with the storage of the common block itself (see Section A.33 on page 260 for examples). Fortran 15 2.9.3.1 default clause 16 Summary 17 18 19 The default clause explicitly determines the data-sharing attributes of variables that are referenced in a parallel or task construct and would otherwise be implicitly determined (see Section 2.9.1.1 on page 84). 20 Syntax 21 The syntax of the default clause is as follows: C/C++ default(shared | none) C/C++ 22 Chapter 2 Directives 93 1 Fortran The syntax of the default clause is as follows: 2 default(private | firstprivate | shared | none) Fortran 3 4 Description 5 6 The default(shared) clause causes all variables referenced in the construct that have implicitly determined data-sharing attributes to be shared. Fortran 7 8 The default(firstprivate) clause causes all variables in the construct that have implicitly determined data-sharing attributes to be firstprivate. 9 10 The default(private) clause causes all variables referenced in the construct that have implicitly determined data-sharing attributes to be private. Fortran 11 12 13 14 The default(none) clause requires that each variable that is referenced in the construct, and that does not have a predetermined data-sharing attribute, must have its data-sharing attribute explicitly determined by being listed in a data-sharing attribute clause. See Section A.30 on page 253 for examples. 15 Restrictions 16 The restrictions to the default clause are as follows: 17 • Only a single default clause may be specified on a parallel or task directive. 18 2.9.3.2 shared clause 19 Summary 20 21 The shared clause declares one or more list items to be shared by tasks generated by a parallel or task construct. 94 OpenMP API • Version 3.1 July 2011 1 Syntax 2 The syntax of the shared clause is as follows: shared(list) 3 Description 4 5 All references to a list item within a task refer to the storage area of the original variable at the point the directive was encountered. 6 7 8 It is the programmer's responsibility to ensure, by adding proper synchronization, that storage shared by an explicit task region does not reach the end of its lifetime before the explicit task region completes its execution. Fortran 9 10 11 12 The association status of a shared pointer becomes undefined upon entry to and on exit from the parallel or task construct if it is associated with a target or a subobject of a target that is in a private, firstprivate, lastprivate, or reduction clause inside the construct. 13 14 15 16 17 Under certain conditions, passing a shared variable to a non-intrinsic procedure may result in the value of the shared variable being copied into temporary storage before the procedure reference, and back out of the temporary storage into the actual argument storage after the procedure reference. It is implementation defined when this situation occurs. See Section A.31 on page 255 for an example of this behavior. 18 19 Note – Use of intervening temporary storage may occur when the following three conditions hold regarding an actual argument in a reference to a non-intrinsic procedure: 20 a. The actual argument is one of the following: 21 • A shared variable. 22 • A subobject of a shared variable. 23 • An object associated with a shared variable. 24 • An object associated with a subobject of a shared variable. 25 b. The actual argument is also one of the following: 26 • An array section. 27 • An array section with a vector subscript. 28 • An assumed-shape array. 29 • A pointer array. Chapter 2 Directives 95 1 2 c. The associated dummy argument for this actual argument is an explicit-shape array or an assumed-size array. 3 4 5 6 These conditions effectively result in references to, and definitions of, the temporary storage during the procedure reference. Any references to (or definitions of) the shared storage that is associated with the dummy argument by any other task must be synchronized with the procedure reference to avoid possible race conditions. 7 Fortran 8 2.9.3.3 private clause Summary 9 10 The private clause declares one or more list items to be private to a task. 11 Syntax 12 The syntax of the private clause is as follows: private(list) 13 Description 14 15 16 17 18 19 20 21 Each task that references a list item that appears in a private clause in any statement in the construct receives a new list item whose language-specific attributes are derived from the original list item. Inside the construct, all references to the original list item are replaced by references to the new list item. In the rest of the region, it is unspecified whether references are to the new list item or the original list item. Therefore, if an attempt is made to reference the original item, its value after the region is also unspecified. If a task does not reference a list item that appears in a private clause, it is unspecified whether that task receives a new list item. 22 The value and/or allocation status of the original list item will change only: 23 • if accessed and modified via pointer, 24 • if possibly accessed in the region but outside of the construct, or 25 • as a side effect of directives or clauses. 96 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 List items that appear in a private, firstprivate, or reduction clause in a parallel construct may also appear in a private clause in an enclosed parallel, task, or worksharing construct. List items that appear in a private or firstprivate clause in a task construct may also appear in a private clause in an enclosed parallel or task construct. List items that appear in a private, firstprivate, lastprivate, or reduction clause in a worksharing construct may also appear in a private clause in an enclosed parallel or task construct. See Section A.32 on page 256 for an example. 9 10 11 12 13 A new list item of the same type, with automatic storage duration, is allocated for the construct. The storage and thus lifetime of these list items lasts until the block in which they are created exits. The size and alignment of the new list item are determined by the type of the variable. This allocation occurs once for each task generated by the construct, if the task references the list item in any statement. 14 15 16 17 The new list item is initialized, or has an undefined initial value, as if it had been locally declared without an initializer. The order in which any default constructors for different private variables of class type are called is unspecified. The order in which any destructors for different private variables of class type are called is unspecified. C/C++ C/C++ Fortran 18 19 20 21 22 A new list item of the same type is allocated once for each implicit task in the parallel region, or for each task generated by a task construct, if the construct references the list item in any statement. The initial value of the new list item is undefined. Within a parallel, worksharing, or task region, the initial status of a private pointer is undefined. 23 For a list item with the ALLOCATABLE attribute: 24 25 • if the list item is "not currently allocated", the new list item will have an initial state 26 27 • if the list item is allocated, the new list item will have an initial state of allocated 28 29 30 31 A list item that appears in a private clause may be storage-associated with other variables when the private clause is encountered. Storage association may exist because of constructs such as EQUIVALENCE or COMMON. If A is a variable appearing in a private clause and B is a variable that is storage-associated with A, then: 32 33 • The contents, allocation, and association status of B are undefined on entry to the 34 35 • Any definition of A, or of its allocation or association status, causes the contents, of "not currently allocated"; with the same array bounds. parallel or task region. allocation, and association status of B to become undefined. Chapter 2 Directives 97 1 2 • Any definition of B, or of its allocation or association status, causes the contents, 3 For examples, see Section A.33 on page 260. allocation, and association status of A to become undefined. Fortran 4 For examples of the private clause, see Section A.32 on page 256. 5 Restrictions 6 The restrictions to the private clause are as follows: 7 8 • A variable that is part of another variable (as an array or structure element) cannot appear in a private clause. C/C++ 9 10 • A variable of class type (or array thereof) that appears in a private clause requires 11 12 13 • A variable that appears in a private clause must not have a const-qualified type 14 15 • A variable that appears in a private clause must not have an incomplete type or a an accessible, unambiguous default constructor for the class type. unless it is of class type with a mutable member. This restriction does not apply to the firstprivate clause. reference type. C/C++ Fortran 16 17 • A variable that appears in a private clause must either be definable, or an 18 19 • Variables that appear in namelist statements, in variable format expressions, and in allocatable array. This restriction does not apply to the firstprivate clause. expressions for statement function definitions, may not appear in a private clause. Fortran 20 2.9.3.4 firstprivate clause 21 Summary 22 23 24 The firstprivate clause declares one or more list items to be private to a task, and initializes each of them with the value that the corresponding original item has when the construct is encountered. 98 OpenMP API • Version 3.1 July 2011 1 Syntax 2 The syntax of the firstprivate clause is as follows: firstprivate(list) 3 Description 4 5 The firstprivate clause provides a superset of the functionality provided by the private clause. 6 7 8 9 10 11 A list item that appears in a firstprivate clause is subject to the private clause semantics described in Section 2.9.3.3 on page 96, except as noted. In addition, the new list item is initialized from the original list item existing before the construct. The initialization of the new list item is done once for each task that references the list item in any statement in the construct. The initialization is done prior to the execution of the construct. 12 13 14 15 16 17 18 For a firstprivate clause on a parallel or task construct, the initial value of the new list item is the value of the original list item that exists immediately prior to the construct in the task region where the construct is encountered. For a firstprivate clause on a worksharing construct, the initial value of the new list item for each implicit task of the threads that execute the worksharing construct is the value of the original list item that exists in the implicit task immediately prior to the point in time that the worksharing construct is encountered. 19 20 21 To avoid race conditions, concurrent updates of the original list item must be synchronized with the read of the original list item that occurs as a result of the firstprivate clause. 22 23 If a list item appears in both firstprivate and lastprivate clauses, the update required for lastprivate occurs after all the initializations for firstprivate. 24 25 26 27 28 29 For variables of non-array type, the initialization occurs by copy assignment. For an array of elements of non-array type, each element is initialized as if by assignment from an element of the original array to the corresponding element of the new array. For variables of class type, a copy constructor is invoked to perform the initialization. The order in which copy constructors for different variables of class type are called is unspecified. C/C++ C/C++ Chapter 2 Directives 99 Fortran If the original list item does not have the POINTER attribute, initialization of the new list items occurs as if by intrinsic assignment, unless the original list item has the allocation status of not currently allocated, in which case the new list items will have the same status. 1 2 3 4 5 6 7 If the original list item has the POINTER attribute, the new list items receive the same association status of the original list item as if by pointer assignment. Fortran 8 Restrictions 9 The restrictions to the firstprivate clause are as follows: 10 11 • A variable that is part of another variable (as an array or structure element) cannot 12 13 14 15 • A list item that is private within a parallel region must not appear in a 16 17 18 19 • A list item that appears in a reduction clause of a parallel construct must not 20 21 22 • A list item that appears in a reduction clause in a worksharing construct must not 23 24 • A variable of class type (or array thereof) that appears in a firstprivate clause 25 26 • A variable that appears in a firstprivate clause must not have an incomplete appear in a firstprivate clause. firstprivate clause on a worksharing construct if any of the worksharing regions arising from the worksharing construct ever bind to any of the parallel regions arising from the parallel construct. appear in a firstprivate clause on a worksharing or task construct if any of the worksharing or task regions arising from the worksharing or task construct ever bind to any of the parallel regions arising from the parallel construct. appear in a firstprivate clause in a task construct encountered during execution of any of the worksharing regions arising from the worksharing construct. C/C++ requires an accessible, unambiguous copy constructor for the class type. type or a reference type. C/C++ 27 Fortran • Variables that appear in namelist statements, in variable format expressions, and in 28 29 30 expressions for statement function definitions, may not appear in a firstprivate clause. Fortran 100 OpenMP API • Version 3.1 July 2011 1 2.9.3.5 lastprivate clause 2 Summary 3 4 5 The lastprivate clause declares one or more list items to be private to an implicit task, and causes the corresponding original list item to be updated after the end of the region. 6 Syntax 7 The syntax of the lastprivate clause is as follows: lastprivate(list) 8 Description 9 10 The lastprivate clause provides a superset of the functionality provided by the private clause. 11 12 13 14 15 A list item that appears in a lastprivate clause is subject to the private clause semantics described in Section 2.9.3.3 on page 96. In addition, when a lastprivate clause appears on the directive that identifies a worksharing construct, the value of each new list item from the sequentially last iteration of the associated loops, or the lexically last section construct, is assigned to the original list item. 16 17 For an array of elements of non-array type, each element is assigned to the corresponding element of the original array. C/C++ C/C++ Fortran 18 19 If the original list item does not have the POINTER attribute, its update occurs as if by intrinsic assignment. 20 21 If the original list item has the POINTER attribute, its update occurs as if by pointer assignment. Fortran 22 23 24 List items that are not assigned a value by the sequentially last iteration of the loops, or by the lexically last section construct, have unspecified values after the construct. Unassigned subcomponents also have unspecified values after the construct. Chapter 2 Directives 101 1 2 3 4 The original list item becomes defined at the end of the construct if there is an implicit barrier at that point. To avoid race conditions, concurrent reads or updates of the original list item must be synchronized with the update of the original list item that occurs as a result of the lastprivate clause. 5 6 7 8 If the lastprivate clause is used on a construct to which nowait is applied, accesses to the original list item may create a data race. To avoid this, synchronization must be inserted to ensure that the sequentially last iteration or lexically last section construct has stored and flushed that list item. 9 10 If a list item appears in both firstprivate and lastprivate clauses, the update required for lastprivate occurs after all initializations for firstprivate. 11 For an example of the lastprivate clause, see Section A.35 on page 264. 12 Restrictions 13 The restrictions to the lastprivate clause are as follows: 14 15 • A variable that is part of another variable (as an array or structure element) cannot 16 17 18 19 • A list item that is private within a parallel region, or that appears in the 20 21 22 • A variable of class type (or array thereof) that appears in a lastprivate clause 23 24 25 26 • A variable of class type (or array thereof) that appears in a lastprivate clause 27 28 • A variable that appears in a lastprivate clause must not have a const-qualified 29 30 • A variable that appears in a lastprivate clause must not have an incomplete type appear in a lastprivate clause. reduction clause of a parallel construct, must not appear in a lastprivate clause on a worksharing construct if any of the corresponding worksharing regions ever binds to any of the corresponding parallel regions. C/C++ requires an accessible, unambiguous default constructor for the class type, unless the list item is also specified in a firstprivate clause. requires an accessible, unambiguous copy assignment operator for the class type. The order in which copy assignment operators for different variables of class type are called is unspecified. type unless it is of class type with a mutable member. or a reference type. C/C++ Fortran • A variable that appears in a lastprivate clause must be definable. 31 102 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 • An original list item with the ALLOCATABLE attribute must be in the allocated state 6 7 8 • Variables that appear in namelist statements, in variable format expressions, and in at entry to the construct containing the lastprivate clause. The list item in the sequentially last iteration or lexically last section must be in the allocated state upon exit from that iteration or section with the same bounds as the corresponding original list item. expressions for statement function definitions, may not appear in a lastprivate clause. Fortran 9 2.9.3.6 reduction clause 10 Summary 11 12 13 14 The reduction clause specifies an operator and one or more list items. For each list item, a private copy is created in each implicit task, and is initialized appropriately for the operator. After the end of the region, the original list item is updated with the values of the private copies using the specified operator. 15 Syntax 16 The syntax of the reduction clause is as follows: C/C++ reduction(operator:list) 17 18 The following table lists the operators that are valid and their initialization values. The actual initialization value depends on the data type of the reduction list item. Operator Initialization value + 0 * 1 - 0 & ~0 | 0 ^ 0 && 1 Chapter 2 Directives 103 || 0 max Least representable value in the reduction list item type min Largest representable value in the reduction list item type 1 C/C++ Fortran The syntax of the reduction clause is as follows: 2 reduction({operator | intrinsic_procedure_name}:list) The following table lists the operators and intrinsic_procedure_names that are valid and their initialization values. The actual initialization value depends on the data type of the reduction list item. 3 4 5 Operator/ Intrinsic Initialization value + 0 * 1 - 0 .and. .true. .or. .false. .eqv. .true. .neqv. .false. max Least representable number in the reduction list item type min Largest representable number in the reduction list item type iand All bits on ior 0 ieor 0 6 Fortran 104 OpenMP API • Version 3.1 July 2011 1 Description 2 3 The reduction clause can be used to perform some forms of recurrence calculations (involving mathematically associative and commutative operators) in parallel. 4 5 6 7 8 9 A private copy of each list item is created, one for each implicit task, as if the private clause had been used. The private copy is then initialized to the initialization value for the operator, as specified above. At the end of the region for which the reduction clause was specified, the original list item is updated by combining its original value with the final value of each of the private copies, using the operator specified. (The partial results of a subtraction reduction are added to form the final value.) 10 11 For max and min operators, the final values of the private copies are combined with the original list item value using the following expressions: C/C++ max original_list_item = original_list_item < private_copy ? private_copy : original_list_item; min original_list_item = original_list_item > private_copy ? private_copy : original_list_item; C/C++ 12 13 14 15 16 17 18 19 If nowait is not used, the reduction computation will be complete at the end of the construct; however, if the reduction clause is used on a construct to which nowait is also applied, accesses to the original list item will create a race and, thus, have unspecified effect unless synchronization ensures that they occur after all threads have executed all of their iterations or section constructs, and the reduction computation has completed and stored the computed value of that list item. This can most simply be ensured through a barrier synchronization. 20 21 22 23 24 25 The location in the OpenMP program at which the values are combined and the order in which the values are combined are unspecified. Therefore, when comparing sequential and parallel runs, or when comparing one parallel run to another (even if the number of threads used is the same), there is no guarantee that bit-identical results will be obtained or that side effects (such as floating point exceptions) will be identical or take place at the same location in the OpenMP program. 26 27 28 To avoid race conditions, concurrent reads or updates of the original list item must be synchronized with the update of the original list item that occurs as a result of the reduction computation. Chapter 2 Directives 105 1 Restrictions 2 The restrictions to the reduction clause are as follows: 3 4 5 • A list item that appears in a reduction clause of a worksharing construct must be 6 7 • A list item that appears in a reduction clause of the innermost enclosing 8 9 • Any number of reduction clauses can be specified on the directive, but a list item shared in the parallel regions to which any of the worksharing regions arising from the worksharing construct bind. worksharing or parallel construct may not be accessed in an explicit task. can appear only once in the reduction clauses for that directive. C/C++ 10 11 12 13 14 15 16 • The type of a list item that appears in a reduction clause must be valid for the 17 18 • Aggregate types (including arrays), pointer types and reference types may not appear 19 • A list item that appears in a reduction clause must not be const-qualified. reduction operator. For a max or min reduction in C, the type of the list item must be an allowed arithmetic data type: char, int, float, double, or _Bool, possibly modified with long, short, signed, or unsigned. For a max or min reduction in C++, the type of the list item must be an allowed arithmetic data type: char, wchar_t, int, float, double, or bool, possibly modified with long, short, signed, or unsigned. in a reduction clause. C/C++ Fortran 20 21 • The type of a list item that appears in a reduction clause must be valid for the 22 • A list item that appears in a reduction clause must be definable. 23 24 • A list item that appears in a reduction clause must be a named variable of 25 26 27 • An original list item with the ALLOCATABLE attribute must be in the allocated state 28 • Fortran pointers and Cray pointers may not appear in a reduction clause. 29 30 31 • Operators specified must be intrinsic operators and any intrinsic_procedure_name reduction operator or intrinsic. intrinsic type. at entry to the construct containing the reduction clause. Additionally, the list item must not be deallocated and/or allocated within the region. must refer to one of the allowed intrinsic procedures. Assignment to the reduction list items must be via intrinsic assignment. See Section A.36 on page 266 for examples. Fortran 106 OpenMP API • Version 3.1 July 2011 1 2.9.4 Data Copying Clauses 2 3 4 This section describes the copyin clause (allowed on the parallel directive and combined parallel worksharing directives) and the copyprivate clause (allowed on the single directive). 5 6 7 These clauses support the copying of data values from private or threadprivate variables on one implicit task or thread to the corresponding variables on other implicit tasks or threads in the team. 8 9 10 11 The clauses accept a comma-separated list of list items (see Section 2.1 on page 22). All list items appearing in a clause must be visible, according to the scoping rules of the base language. Clauses may be repeated as needed, but a list item that specifies a given variable may not appear in more than one clause on the same directive. 12 2.9.4.1 copyin clause 13 Summary 14 15 16 The copyin clause provides a mechanism to copy the value of the master thread’s threadprivate variable to the threadprivate variable of each other member of the team executing the parallel region. 17 Syntax 18 The syntax of the copyin clause is as follows: copyin(list) 19 Description 20 21 22 23 24 25 26 The copy is done after the team is formed and prior to the start of execution of the associated structured block. For variables of non-array type, the copy occurs by copy assignment. For an array of elements of non-array type, each element is copied as if by assignment from an element of the master thread’s array to the corresponding element of the other thread’s array. For class types, the copy assignment operator is invoked. The order in which copy assignment operators for different variables of class type are called is unspecified. C/C++ C/C++ Chapter 2 Directives 107 Fortran 1 2 The copy is done, as if by assignment, after the team is formed and prior to the start of execution of the associated structured block. 3 4 5 On entry to any parallel region, each thread’s copy of a variable that is affected by a copyin clause for the parallel region will acquire the allocation, association, and definition status of the master thread’s copy, according to the following rules: 6 7 • If the original list item has the POINTER attribute, each copy receives the same association status of the master thread’s copy as if by pointer assignment. • If the original list item does not have the POINTER attribute, each copy becomes 8 9 10 11 defined with the value of the master thread's copy as if by intrinsic assignment, unless it has the allocation status of not currently allocated, in which case each copy will have the same status. Fortran 12 For an example of the copyin clause, see Section A.37 on page 271. 13 Restrictions 14 The restrictions to the copyin clause are as follows: 15 • A list item that appears in a copyin clause must be threadprivate. 16 17 • A variable of class type (or array thereof) that appears in a copyin clause requires C/C++ an accessible, unambiguous copy assignment operator for the class type. C/C++ Fortran 18 19 20 • A list item that appears in a copyin clause must be threadprivate. Named variables 21 22 • A common block name that appears in a copyin clause must be declared to be a 23 24 • If an array with the ALLOCATABLE attribute is allocated, then each thread's copy of appearing in a threadprivate common block may be specified: it is not necessary to specify the whole common block. common block in the same scoping unit in which the copyin clause appears. that array must be allocated with the same bounds. Fortran 108 OpenMP API • Version 3.1 July 2011 1 2.9.4.2 copyprivate clause 2 Summary 3 4 5 The copyprivate clause provides a mechanism to use a private variable to broadcast a value from the data environment of one implicit task to the data environments of the other implicit tasks belonging to the parallel region. 6 7 8 To avoid race conditions, concurrent reads or updates of the list item must be synchronized with the update of the list item that occurs as a result of the copyprivate clause. 9 Syntax 10 The syntax of the copyprivate clause is as follows: copyprivate(list) 11 Description 12 13 14 15 The effect of the copyprivate clause on the specified list items occurs after the execution of the structured block associated with the single construct (see Section 2.5.3 on page 50), and before any of the threads in the team have left the barrier at the end of the construct. 16 17 18 19 20 21 22 23 24 In all other implicit tasks belonging to the parallel region, each specified list item becomes defined with the value of the corresponding list item in the implicit task whose thread executed the structured block. For variables of non-array type, the definition occurs by copy assignment. For an array of elements of non-array type, each element is copied by copy assignment from an element of the array in the data environment of the implicit task associated with the thread that executed the structured block to the corresponding element of the array in the data environment of the other implicit tasks. For class types, a copy assignment operator is invoked. The order in which copy assignment operators for different variables of class type are called is unspecified. C/C++ C/C++ Fortran 25 26 27 28 If a list item does not have the POINTER attribute, then in all other implicit tasks belonging to the parallel region, the list item becomes defined as if by intrinsic assignment with the value of the corresponding list item in the implicit task associated with the thread that executed the structured block. Chapter 2 Directives 109 If the list item has the POINTER attribute, then, in all other implicit tasks belonging to the parallel region, the list item receives, as if by pointer assignment, the same association status of the corresponding list item in the implicit task associated with the thread that executed the structured block. 1 2 3 4 Fortran 5 For examples of the copyprivate clause, see Section A.38 on page 273. 6 7 8 Note – The copyprivate clause is an alternative to using a shared variable for the value when providing such a shared variable would be difficult (for example, in a recursion requiring a different variable at each level). 9 Restrictions 10 The restrictions to the copyprivate clause are as follows: 11 12 • All list items that appear in the copyprivate clause must be either threadprivate 13 14 • A list item that appears in a copyprivate clause may not appear in a private or 15 16 • A variable of class type (or array thereof) that appears in a copyprivate clause or private in the enclosing context. firstprivate clause on the single construct. C/C++ requires an accessible unambiguous copy assignment operator for the class type. C/C++ Fortran 17 • A common block that appears in a copyprivate clause must be threadprivate. 18 19 • An array with the ALLOCATABLE attribute must be in the allocated state with the same bounds in all threads affected by the copyprivate clause. Fortran 110 OpenMP API • Version 3.1 July 2011 1 2.10 Nesting of Regions 2 3 This section describes a set of restrictions on the nesting of regions. The restrictions on nesting are as follows: 4 5 • A worksharing region may not be closely nested inside a worksharing, explicit task, 6 7 • A barrier region may not be closely nested inside a worksharing, explicit task, 8 9 • A master region may not be closely nested inside a worksharing, atomic, or critical, ordered, atomic, or master region. critical, ordered, atomic, or master region. explicit task region. 10 11 • An ordered region may not be closely nested inside a critical, atomic, or 12 13 • An ordered region must be closely nested inside a loop region (or parallel loop 14 15 16 • A critical region may not be nested (closely or otherwise) inside a critical 17 18 • parallel, flush, critical, atomic, taskyield, and explicit task 19 20 For examples illustrating these rules, see Section A.20 on page 221, Section A.39 on page 278, Section A.40 on page 281, and Section A.15 on page 193. explicit task region. region) with an ordered clause. region with the same name. Note that this restriction is not sufficient to prevent deadlock. regions may not be closely nested inside an atomic region. Chapter 2 Directives 111 1 This page intentionally left blank. 2 3 112 OpenMP API • Version 3.1 July 2011 1 CHAPTER 3 2 Runtime Library Routines 3 4 This chapter describes the OpenMP API runtime library routines and is divided into the following sections: 5 • Runtime library definitions (Section 3.1 on page 114). 6 7 • Execution environment routines that can be used to control and to query the parallel 8 9 • Lock routines that can be used to synchronize access to data (Section 3.3 on page 10 execution environment (Section 3.2 on page 115). 141). • Portable timer routines (Section 3.4 on page 148). 11 12 13 Throughout this chapter, true and false are used as generic terms to simplify the description of the routines. 14 true means a nonzero integer value and false means an integer value of zero. C/C++ C/C++ Fortran 15 true means a logical value of .TRUE. and false means a logical value of .FALSE.. Fortran Fortran 16 Restrictions 17 The following restriction applies to all OpenMP runtime library routines: 18 19 • OpenMP runtime library routines may not be called from PURE or ELEMENTAL procedures. Fortran 113 1 3.1 Runtime Library Definitions 2 3 4 5 6 For each base language, a compliant implementation must supply a set of definitions for the OpenMP API runtime library routines and the special data types of their parameters. The set of definitions must contain a declaration for each OpenMP API runtime library routine and a declaration for the simple lock, nestable lock and schedule data types. In addition, each set of definitions may specify other implementation specific values. 7 The library routines are external functions with “C” linkage. 8 9 Prototypes for the C/C++ runtime library routines described in this chapter shall be provided in a header file named omp.h. This file defines the following: C/C++ 10 • The prototypes of all the routines in the chapter. 11 • The type omp_lock_t. 12 • The type omp_nest_lock_t. 13 • The type omp_sched_t. 14 See Section D.1 on page 326 for an example of this file. C/C++ Fortran 15 16 The OpenMP Fortran API runtime library routines are external procedures. The return values of these routines are of default kind, unless otherwise specified. 17 18 19 20 Interface declarations for the OpenMP Fortran runtime library routines described in this chapter shall be provided in the form of a Fortran include file named omp_lib.h or a Fortran 90 module named omp_lib. It is implementation defined whether the include file or the module file (or both) is provided. 21 These files define the following: 22 • The interfaces of all of the routines in this chapter. 23 • The integer parameter omp_lock_kind. 24 • The integer parameter omp_nest_lock_kind. 25 • The integer parameter omp_sched_kind. 26 27 28 29 30 • The integer parameter openmp_version with a value yyyymm where yyyy and mm are the year and month designations of the version of the OpenMP Fortran API that the implementation supports. This value matches that of the C preprocessor macro _OPENMP, when a macro preprocessor is supported (see Section 2.2 on page 26). 114 OpenMP API • Version 3.1 July 2011 1 See Section D.2 on page 328 and Section D.3 on page 330 for examples of these files. 2 3 4 It is implementation defined whether any of the OpenMP runtime library routines that take an argument are extended with a generic interface so arguments of different KIND type can be accommodated. See Appendix D.4 for an example of such an extension. Fortran 5 3.2 Execution Environment Routines 6 7 The routines described in this section affect and monitor threads, processors, and the parallel environment. 8 • the omp_set_num_threads routine. 9 • the omp_get_num_threads routine. 10 • the omp_get_max_threads routine. 11 • the omp_get_thread_num routine. 12 • the omp_get_num_procs routine. 13 • the omp_in_parallel routine. 14 • the omp_set_dynamic routine. 15 • the omp_get_dynamic routine. 16 • the omp_set_nested routine. 17 • the omp_get_nested routine. 18 • the omp_set_schedule routine. 19 • the omp_get_schedule routine. 20 • the omp_get_thread_limit routine. 21 • the omp_set_max_active_levels routine. 22 • the omp_get_max_active_levels routine. 23 • the omp_get_level routine. 24 • the omp_get_ancestor_thread_num routine. 25 • the omp_get_team_size routine. 26 • the omp_get_active_level routine. 27 • the omp_in_final routine. Chapter 3 Runtime Library Routines 115 1 3.2.1 omp_set_num_threads 2 Summary 3 4 5 The omp_set_num_threads routine affects the number of threads to be used for subsequent parallel regions that do not specify a num_threads clause, by setting the value of the first element of the nthreads-var ICV of the current task. 6 Format C/C++ void omp_set_num_threads(int num_threads); C/C++ 7 Fortran subroutine omp_set_num_threads(num_threads) integer num_threads Fortran 8 Constraints on Arguments 9 10 11 The value of the argument passed to this routine must evaluate to a positive integer, or else the behavior of this routine is implementation defined. 12 Binding 13 The binding task set for an omp_set_num_threads region is the generating task. 14 Effect 15 16 The effect of this routine is to set the value of the first element of the nthreads-var ICV of the current task to the value specified in the argument. 17 18 See Section 2.4.1 on page 36 for the rules governing the number of threads used to execute a parallel region. 116 OpenMP API • Version 3.1 July 2011 1 2 For an example of the omp_set_num_threads routine, see Section A.41 on page 288. 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 155. 6 • omp_get_max_threads routine, see Section 3.2.3 on page 118. 7 • parallel construct, see Section 2.4 on page 33. 8 • num_threads clause, see Section 2.4 on page 33. 9 3.2.2 omp_get_num_threads 10 Summary 11 12 The omp_get_num_threads routine returns the number of threads in the current team. 13 Format C/C++ int omp_get_num_threads(void); C/C++ 14 Fortran integer function omp_get_num_threads() Fortran 15 16 Binding 17 18 The binding region for an omp_get_num_threads region is the innermost enclosing parallel region. Chapter 3 Runtime Library Routines 117 1 Effect 2 3 4 5 The omp_get_num_threads routine returns the number of threads in the team executing the parallel region to which the routine region binds. If called from the sequential part of a program, this routine returns 1. For examples, see Section A.42 on page 289. 6 7 See Section 2.4.1 on page 36 for the rules governing the number of threads used to execute a parallel region. 8 Cross References 9 • parallel construct, see Section 2.4 on page 33. 10 • omp_set_num_threads routine, see Section 3.2.1 on page 116. 11 • OMP_NUM_THREADS environment variable, see Section 4.2 on page 155. 12 3.2.3 omp_get_max_threads 13 Summary 14 15 16 The omp_get_max_threads routine returns an upper bound on the number of threads that could be used to form a new team if a parallel region without a num_threads clause were encountered after execution returns from this routine. 17 Format C/C++ int omp_get_max_threads(void); C/C++ 18 Fortran integer function omp_get_max_threads() Fortran 19 118 OpenMP API • Version 3.1 July 2011 1 Binding 2 The binding task set for an omp_get_max_threads region is the generating task. 3 Effect 4 5 6 7 The value returned by omp_get_max_threads is the value of the first element of the nthreads-var ICV of the current task. This value is also an upper bound on the number of threads that could be used to form a new team if a parallel region without a num_threads clause were encountered after execution returns from this routine. 8 9 See Section 2.4.1 on page 36 for the rules governing the number of threads used to execute a parallel region. 10 11 12 Note – The return value of the omp_get_max_threads routine can be used to dynamically allocate sufficient storage for all threads in the team formed at the subsequent active parallel region. 13 Cross References 14 • nthreads-var ICV, see Section 2.3 on page 28. 15 • parallel construct, see Section 2.4 on page 33. 16 • num_threads clause, see Section 2.4 on page 33. 17 • omp_set_num_threads routine, see Section 3.2.1 on page 116. 18 • OMP_NUM_THREADS environment variable, see Section 4.2 on page 155. 19 3.2.4 omp_get_thread_num 20 Summary 21 22 The omp_get_thread_num routine returns the thread number, within the current team, of the calling thread. Chapter 3 Runtime Library Routines 119 Format 1 C/C++ int omp_get_thread_num(void); C/C++ 2 Fortran integer function omp_get_thread_num() Fortran 3 4 Binding 5 6 7 The binding thread set for an omp_get_thread_num region is the current team. The binding region for an omp_get_thread_num region is the innermost enclosing parallel region. 8 Effect 9 10 11 12 13 The omp_get_thread_num routine returns the thread number of the calling thread, within the team executing the parallel region to which the routine region binds. The thread number is an integer between 0 and one less than the value returned by omp_get_num_threads, inclusive. The thread number of the master thread of the team is 0. The routine returns 0 if it is called from the sequential part of a program. 14 15 16 Note – The thread number may change at any time during the execution of an untied task. The value returned by omp_get_thread_num is not generally useful during the execution of such a task region. 17 Cross References 18 • omp_get_num_threads routine, see Section 3.2.2 on page 117. 120 OpenMP API • Version 3.1 July 2011 1 3.2.5 omp_get_num_procs 2 Summary 3 4 The omp_get_num_procs routine returns the number of processors available to the program. 5 Format C/C++ int omp_get_num_procs(void); C/C++ 6 Fortran integer function omp_get_num_procs() Fortran 7 8 Binding 9 10 11 The binding thread set for an omp_get_num_procs region is all threads. The effect of executing this routine is not related to any specific region corresponding to any construct or API routine. 12 Effect 13 14 15 16 17 The omp_get_num_procs routine returns the number of processors that are available to the program at the time the routine is called. Note that this value may change between the time that it is determined by the omp_get_num_procs routine and the time that it is read in the calling context due to system actions outside the control of the OpenMP implementation. Chapter 3 Runtime Library Routines 121 1 3.2.6 omp_in_parallel 2 Summary 3 4 The omp_in_parallel routine returns true if the call to the routine is enclosed by an active parallel region; otherwise, it returns false. 5 Format C/C++ int omp_in_parallel(void); C/C++ 6 Fortran logical function omp_in_parallel() Fortran 7 Binding 8 9 10 11 The binding thread set for an omp_in_parallel region is all threads. The effect of executing this routine is not related to any specific parallel region but instead depends on the state of all enclosing parallel regions. 12 Effect 13 14 15 omp_in_parallel returns true if any enclosing parallel region is active. If the routine call is enclosed by only inactive parallel regions (including the implicit parallel region), then it returns false. 122 OpenMP API • Version 3.1 July 2011 1 3.2.7 omp_set_dynamic 2 Summary 3 4 5 The omp_set_dynamic routine enables or disables dynamic adjustment of the number of threads available for the execution of subsequent parallel regions by setting the value of the dyn-var ICV. 6 Format C/C++ void omp_set_dynamic(int dynamic_threads); C/C++ 7 Fortran subroutine omp_set_dynamic (dynamic_threads) logical dynamic_threads Fortran 8 9 Binding 10 The binding task set for an omp_set_dynamic region is the generating task. 11 Effect 12 13 14 15 16 For implementations that support dynamic adjustment of the number of threads, if the argument to omp_set_dynamic evaluates to true, dynamic adjustment is enabled for the current task; otherwise, dynamic adjustment is disabled for the current task. For implementations that do not support dynamic adjustment of the number of threads this routine has no effect: the value of dyn-var remains false. 17 For an example of the omp_set_dynamic routine, see Section A.41 on page 288. 18 19 See Section 2.4.1 on page 36 for the rules governing the number of threads used to execute a parallel region. Chapter 3 Runtime Library Routines 123 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 117. 4 • omp_get_dynamic routine, see Section 3.2.8 on page 124. 5 • OMP_DYNAMIC environment variable, see Section 4.3 on page 156. 6 3.2.8 omp_get_dynamic 7 Summary 8 9 The omp_get_dynamic routine returns the value of the dyn-var ICV, which determines whether dynamic adjustment of the number of threads is enabled or disabled. Format 10 C/C++ int omp_get_dynamic(void); C/C++ 11 Fortran logical function omp_get_dynamic() Fortran 12 13 Binding 14 The binding task set for an omp_get_dynamic region is the generating task. 15 Effect 16 17 18 This routine returns true if dynamic adjustment of the number of threads is enabled for the current task; it returns false, otherwise. If an implementation does not support dynamic adjustment of the number of threads, then this routine always returns false. 124 OpenMP API • Version 3.1 July 2011 1 2 See Section 2.4.1 on page 36 for the rules governing the number of threads used to execute a parallel region. 3 Cross References 4 • dyn-var ICV, see Section 2.3 on page 28. 5 • omp_set_dynamic routine, see Section 3.2.7 on page 123. 6 • OMP_DYNAMIC environment variable, see Section 4.3 on page 156. 7 8 3.2.9 omp_set_nested Summary 9 10 The omp_set_nested routine enables or disables nested parallelism, by setting the nest-var ICV. 11 Format C/C++ void omp_set_nested(int nested); 12 C/C++ Fortran subroutine omp_set_nested (nested) logical nested 13 Fortran Chapter 3 Runtime Library Routines 125 1 Binding 2 The binding task set for an omp_set_nested region is the generating task. 3 Effect 4 5 6 7 8 For implementations that support nested parallelism, if the argument to omp_set_nested evaluates to true, nested parallelism is enabled for the current task; otherwise, nested parallelism is disabled for the current task. For implementations that do not support nested parallelism, this routine has no effect: the value of nest-var remains false. 9 10 See Section 2.4.1 on page 36 for the rules governing the number of threads used to execute a parallel region. 11 Cross References 12 • nest-var ICV, see Section 2.3 on page 28. 13 • omp_set_max_active_levels routine, see Section 3.2.14 on page 132. 14 • omp_get_max_active_levels routine, see Section 3.2.15 on page 134. 15 • omp_get_nested routine, see Section 3.2.10 on page 126. 16 • OMP_NESTED environment variable, see Section 4.5 on page 157. 17 3.2.10 omp_get_nested 18 Summary 19 20 The omp_get_nested routine returns the value of the nest-var ICV, which determines if nested parallelism is enabled or disabled. 126 OpenMP API • Version 3.1 July 2011 1 Format C/C++ int omp_get_nested(void); C/C++ 2 Fortran logical function omp_get_nested() Fortran 3 4 Binding 5 The binding task set for an omp_get_nested region is the generating task. 6 Effect 7 8 9 This routine returns true if nested parallelism is enabled for the current task; it returns false, otherwise. If an implementation does not support nested parallelism, this routine always returns false. 10 11 See Section 2.4.1 on page 36 for the rules governing the number of threads used to execute a parallel region. 12 Cross References 13 • nest-var ICV, see Section 2.3 on page 28. 14 • omp_set_nested routine, see Section 3.2.9 on page 125. 15 • OMP_NESTED environment variable, see Section 4.5 on page 157. Chapter 3 Runtime Library Routines 127 1 3.2.11 omp_set_schedule 2 Summary 3 4 The omp_set_schedule routine affects the schedule that is applied when runtime is used as schedule kind, by setting the value of the run-sched-var ICV. 5 Format 6 C/C++ void omp_set_schedule(omp_sched_t kind, int modifier); 7 C/C++ 8 Fortran subroutine omp_set_schedule(kind, modifier) integer (kind=omp_sched_kind) kind integer modifier Fortran 9 10 Constraints on Arguments 11 12 13 14 15 The first argument passed to this routine can be one of the valid OpenMP schedule kinds (except for runtime) or any implementation specific schedule. The C/C++ header file (omp.h) and the Fortran include file (omp_lib.h) and/or Fortran 90 module file (omp_lib) define the valid constants. The valid constants must include the following, which can be extended with implementation specific values: 128 OpenMP API • Version 3.1 July 2011 C/C++ 1 typedef enum omp_sched_t { omp_sched_static = 1, omp_sched_dynamic = 2, omp_sched_guided = 3, omp_sched_auto = 4 } omp_sched_t; C/C++ 2 Fortran integer(kind=omp_sched_kind), integer(kind=omp_sched_kind), integer(kind=omp_sched_kind), integer(kind=omp_sched_kind), parameter parameter parameter parameter :: :: :: :: omp_sched_static = 1 omp_sched_dynamic = 2 omp_sched_guided = 3 omp_sched_auto = 4 Fortran 3 4 Binding 5 The binding task set for an omp_set_schedule region is the generating task. 6 Effect 7 8 9 10 11 12 13 14 The effect of this routine is to set the value of the run-sched-var ICV of the current task to the values specified in the two arguments. The schedule is set to the schedule type specified by the first argument kind. It can be any of the standard schedule types or any other implementation specific one. For the schedule types static, dynamic, and guided the chunk_size is set to the value of the second argument, or to the default chunk_size if the value of the second argument is less than 1; for the schedule type auto the second argument has no meaning; for implementation specific schedule types, the values and associated meanings of the second argument are implementation defined. Chapter 3 Runtime Library Routines 129 1 Cross References 2 • run-sched-var ICV, see Section 2.3 on page 28. 3 • omp_get_schedule routine, see Section 3.2.12 on page 130. 4 • OMP_SCHEDULE environment variable, see Section 4.1 on page 154. 5 • Determining the schedule of a worksharing loop, see Section 2.5.1.1 on page 47. 6 3.2.12 omp_get_schedule 7 Summary 8 9 The omp_get_schedule routine returns the schedule that is applied when the runtime schedule is used. Format 10 11 C/C++ void omp_get_schedule(omp_sched_t * kind, int * modifier ); C/C++ 12 Fortran subroutine omp_get_schedule(kind, modifier) integer (kind=omp_sched_kind) kind integer modifier Fortran 13 14 Binding 15 The binding task set for an omp_get_schedule region is the generating task. 130 OpenMP API • Version 3.1 July 2011 1 Effect 2 3 4 5 6 This routine returns the run-sched-var ICV in the task to which the routine binds. The first argument kind returns the schedule to be used. It can be any of the standard schedule types as defined in Section 3.2.11 on page 128, or any implementation specific schedule type. The second argument is interpreted as in the omp_set_schedule call, defined in Section 3.2.11 on page 128. 7 Cross References 8 • run-sched-var ICV, see Section 2.3 on page 28. 9 • omp_set_schedule routine, see Section 3.2.11 on page 128. 10 • OMP_SCHEDULE environment variable, see Section 4.1 on page 154. 11 • Determining the schedule of a worksharing loop, see Section 2.5.1.1 on page 47. 12 3.2.13 omp_get_thread_limit 13 Summary 14 15 The omp_get_thread_limit routine returns the maximum number of OpenMP threads available to the program. 16 Format 17 C/C++ int omp_get_thread_limit(void); C/C++ 18 Fortran integer function omp_get_thread_limit() 19 Fortran Chapter 3 Runtime Library Routines 131 1 Binding 2 3 4 The binding thread set for an omp_get_thread_limit region is all threads. The effect of executing this routine is not related to any specific region corresponding to any construct or API routine. 5 Effect 6 7 The omp_get_thread_limit routine returns the maximum number of OpenMP threads available to the program as stored in the ICV thread-limit-var. 8 Cross References 9 • thread-limit-var ICV, see Section 2.3 on page 28. • OMP_THREAD_LIMIT environment variable, see Section 4.9 on page 160. 10 11 3.2.14 omp_set_max_active_levels 12 Summary 13 14 The omp_set_max_active_levels routine limits the number of nested active parallel regions, by setting the max-active-levels-var ICV. 15 Format 16 C/C++ void omp_set_max_active_levels (int max_levels); C/C++ 17 132 OpenMP API • Version 3.1 July 2011 1 Fortran subroutine omp_set_max_active_levels (max_levels) integer max_levels Fortran 2 3 Constraints on Arguments 4 5 The value of the argument passed to this routine must evaluate to a non-negative integer, otherwise the behavior of this routine is implementation defined. 6 Binding 7 8 9 10 When called from the sequential part of the program, the binding thread set for an omp_set_max_active_levels region is the encountering thread. When called from within any explicit parallel region, the binding thread set (and binding region, if required) for the omp_set_max_active_levels region is implementation defined. 11 Effect 12 13 The effect of this routine is to set the value of the max-active-levels-var ICV to the value specified in the argument. 14 15 16 If the number of parallel levels requested exceeds the number of levels of parallelism supported by the implementation, the value of the max-active-levels-var ICV will be set to the number of parallel levels supported by the implementation. 17 18 19 This routine has the described effect only when called from the sequential part of the program. When called from within an explicit parallel region, the effect of this routine is implementation defined. 20 Cross References 21 • max-active-levels-var ICV, see Section 2.3 on page 28. 22 • omp_get_max_active_levels routine, see Section 3.2.15 on page 134. 23 • OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.8 on page 159. Chapter 3 Runtime Library Routines 133 1 3.2.15 omp_get_max_active_levels 2 Summary 3 4 5 The omp_get_max_active_levels routine returns the value of the max-activelevels-var ICV, which determines the maximum number of nested active parallel regions. 6 Format 7 C/C++ int omp_get_max_active_levels(void); C/C++ 8 Fortran integer function omp_get_max_active_levels() Fortran 9 10 Binding 11 12 13 14 When called from the sequential part of the program, the binding thread set for an omp_get_max_active_levels region is the encountering thread. When called from within any explicit parallel region, the binding thread set (and binding region, if required) for the omp_get_max_active_levels region is implementation defined. 15 Effect 16 17 18 The omp_get_max_active_levels routine returns the value of the max-activelevels-var ICV, which determines the maximum number of nested active parallel regions. 134 OpenMP API • Version 3.1 July 2011 1 Cross References 2 • max-active-levels-var ICV, see Section 2.3 on page 28. 3 • omp_set_max_active_levels routine, see Section 3.2.14 on page 132. 4 • OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.8 on page 159. 5 3.2.16 omp_get_level 6 Summary 7 8 The omp_get_level routine returns the number of nested parallel regions enclosing the task that contains the call. 9 Format 10 C/C++ int omp_get_level(void); C/C++ 11 Fortran integer function omp_get_level() Fortran 12 13 Binding 14 15 16 The binding task set for an omp_get_level region is the generating task. The binding region for an omp_get_level region is the innermost enclosing parallel region. Chapter 3 Runtime Library Routines 135 1 Effect 2 3 4 5 The omp_get_level routine returns the number of nested parallel regions (whether active or inactive) enclosing the task that contains the call, not including the implicit parallel region. The routine always returns a non-negative integer, and returns 0 if it is called from the sequential part of the program. 6 Cross References 7 • omp_get_active_level routine, see Section 3.2.19 on page 139. 8 • OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.8 on page 159. 9 3.2.17 omp_get_ancestor_thread_num 10 Summary 11 12 The omp_get_ancestor_thread_num routine returns, for a given nested level of the current thread, the thread number of the ancestor or the current thread. 13 Format 14 C/C++ int omp_get_ancestor_thread_num(int level); C/C++ 15 Fortran integer function omp_get_ancestor_thread_num(level) integer level Fortran 16 136 OpenMP API • Version 3.1 July 2011 1 Binding 2 3 4 The binding thread set for an omp_get_ancestor_thread_num region is the encountering thread. The binding region for an omp_get_ancestor_thread_num region is the innermost enclosing parallel region. 5 Effect 6 7 8 9 The omp_get_ancestor_thread_num routine returns the thread number of the ancestor at a given nest level of the current thread or the thread number of the current thread. If the requested nest level is outside the range of 0 and the nest level of the current thread, as returned by the omp_get_level routine, the routine returns -1. 10 11 12 Note – When the omp_get_ancestor_thread_num routine is called with a value of level=0, the routine always returns 0. If level=omp_get_level(), the routine has the same effect as the omp_get_thread_num routine. 13 Cross References 14 • omp_get_level routine, see Section 3.2.16 on page 135. 15 • omp_get_thread_num routine, see Section 3.2.4 on page 119. 16 • omp_get_team_size routine, see Section 3.2.18 on page 137. 17 3.2.18 omp_get_team_size 18 Summary 19 20 The omp_get_team_size routine returns, for a given nested level of the current thread, the size of the thread team to which the ancestor or the current thread belongs. Chapter 3 Runtime Library Routines 137 Format 1 2 C/C++ int omp_get_team_size(int level); C/C++ 3 Fortran integer function omp_get_team_size(level) integer level Fortran 4 5 Binding 6 7 8 The binding thread set for an omp_get_team_size region is the encountering thread. The binding region for an omp_get_team_size region is the innermost enclosing parallel region. 9 Effect 10 11 12 13 14 The omp_get_team_size routine returns the size of the thread team to which the ancestor or the current thread belongs. If the requested nested level is outside the range of 0 and the nested level of the current thread, as returned by the omp_get_level routine, the routine returns -1. Inactive parallel regions are regarded like active parallel regions executed with one thread. 15 16 17 Note – When the omp_get_team_size routine is called with a value of level=0, the routine always returns 1. If level=omp_get_level(), the routine has the same effect as the omp_get_num_threads routine. 138 OpenMP API • Version 3.1 July 2011 1 Cross References 2 • omp_get_num_threads routine, see Section 3.2.2 on page 117. 3 • omp_get_level routine, see Section 3.2.16 on page 135. 4 • omp_get_ancestor_thread_num routine, see Section 3.2.17 on page 136. 5 3.2.19 omp_get_active_level 6 Summary 7 8 The omp_get_active_level routine returns the number of nested, active parallel regions enclosing the task that contains the call. 9 Format 10 C/C++ int omp_get_active_level(void); C/C++ 11 Fortran integer function omp_get_active_level() Fortran 12 13 Binding 14 15 16 The binding task set for the an omp_get_active_level region is the generating task. The binding region for an omp_get_active_level region is the innermost enclosing parallel region. Chapter 3 Runtime Library Routines 139 1 Effect 2 3 4 The omp_get_active_level routine returns the number of nested, active parallel regions enclosing the task that contains the call. The routine always returns a nonnegative integer, and returns 0 if it is called from the sequential part of the program. 5 Cross References 6 • omp_get_level routine, see Section 3.2.16 on page 135. 7 3.2.20 omp_in_final Summary 8 9 10 The omp_in_final routine returns true if the routine is executed in a final task region; otherwise, it returns false. 11 Format 12 C/C++ int omp_in_final(void); C/C++ 13 Fortran logical function omp_in_final() Fortran 14 15 Binding 16 The binding task set for an omp_in_final region is the generating task. 140 OpenMP API • Version 3.1 July 2011 1 Effect 2 3 omp_in_final returns true if the enclosing task region is final. Otherwise, it returns false. 4 3.3 Lock Routines 5 6 7 8 9 The OpenMP runtime library includes a set of general-purpose lock routines that can be used for synchronization. These general-purpose lock routines operate on OpenMP locks that are represented by OpenMP lock variables. OpenMP lock variables must be accessed only through the routines described in this section; programs that otherwise access OpenMP lock variables are non-conforming. 10 11 12 13 14 An OpenMP lock can be in one of the following states: uninitialized, unlocked, or locked. If a lock is in the unlocked state, a task can set the lock, which changes its state to locked. The task that sets the lock is then said to own the lock. A task that owns a lock can unset that lock, returning it to the unlocked state. A program in which a task unsets a lock that is owned by another task is non-conforming. 15 16 17 18 19 Two types of locks are supported: simple locks and nestable locks. A nestable lock can be set multiple times by the same task before being unset; a simple lock cannot be set if it is already owned by the task trying to set it. Simple lock variables are associated with simple locks and can only be passed to simple lock routines. Nestable lock variables are associated with nestable locks and can only be passed to nestable lock routines. 20 21 22 Constraints on the state and ownership of the lock accessed by each of the lock routines are described with the routine. If these constraints are not met, the behavior of the routine is unspecified. 23 24 25 26 The OpenMP lock routines access a lock variable in such a way that they always read and update the most current value of the lock variable. It is not necessary for an OpenMP program to include explicit flush directives to ensure that the lock variable’s value is consistent among different tasks. 27 28 See Section A.45 on page 294 and Section A.46 on page 297, for examples of using the simple and the nestable lock routines, respectively. 29 Binding 30 31 32 The binding thread set for all lock routine regions is all threads. As a consequence, for each OpenMP lock, the lock routine effects relate to all tasks that call the routines, without regard to which teams the threads executing the tasks belong. Chapter 3 Runtime Library Routines 141 1 Simple Lock Routines 2 3 4 The type omp_lock_t is a data type capable of representing a simple lock. For the following routines, a simple lock variable must be of omp_lock_t type. All simple lock routines require an argument that is a pointer to a variable of type omp_lock_t. C/C++ C/C++ Fortran For the following routines, a simple lock variable must be an integer variable of kind=omp_lock_kind. 5 6 Fortran 7 The simple lock routines are as follows: 8 • The omp_init_lock routine initializes a simple lock. 9 • The omp_destroy_lock routine uninitializes a simple lock. 10 • The omp_set_lock routine waits until a simple lock is available, and then sets it. 11 • The omp_unset_lock routine unsets a simple lock. 12 • The omp_test_lock routine tests a simple lock, and sets it if it is available. 13 14 Nestable Lock Routines: 15 16 17 18 The type omp_nest_lock_t is a data type capable of representing a nestable lock. For the following routines, a nested lock variable must be of omp_nest_lock_t type. All nestable lock routines require an argument that is a pointer to a variable of type omp_nest_lock_t. C/C++ C/C++ Fortran For the following routines, a nested lock variable must be an integer variable of kind=omp_nest_lock_kind. 19 20 Fortran 21 The nestable lock routines are as follows: 22 • The omp_init_nest_lock routine initializes a nestable lock. 23 • The omp_destroy_nest_lock routine uninitializes a nestable lock. 142 OpenMP API • Version 3.1 July 2011 1 2 • The omp_set_nest_lock routine waits until a nestable lock is available, and then 3 • The omp_unset_nest_lock routine unsets a nestable lock. 4 5 • The omp_test_nest_lock routine tests a nestable lock, and sets it if it is 6 sets it. available. 3.3.1 omp_init_lock and omp_init_nest_lock 7 Summary 8 These routines provide the only means of initializing an OpenMP lock. 9 Format C/C++ void omp_init_lock(omp_lock_t *lock); void omp_init_nest_lock(omp_nest_lock_t *lock); C/C++ 10 Fortran subroutine omp_init_lock(svar) integer (kind=omp_lock_kind) svar subroutine omp_init_nest_lock(nvar) integer (kind=omp_nest_lock_kind) nvar Fortran 11 12 Constraints on Arguments 13 14 A program that accesses a lock that is not in the uninitialized state through either routine is non-conforming. Chapter 3 Runtime Library Routines 143 1 Effect 2 3 The effect of these routines is to initialize the lock to the unlocked state; that is, no task owns the lock. In addition, the nesting count for a nestable lock is set to zero. 4 For an example of the omp_init_lock routine, see Section A.43 on page 292. 6 omp_destroy_lock and omp_destroy_nest_lock 7 Summary 8 These routines ensure that the OpenMP lock is uninitialized. 9 Format 5 3.3.2 C/C++ void omp_destroy_lock(omp_lock_t *lock); void omp_destroy_nest_lock(omp_nest_lock_t *lock); C/C++ 10 Fortran subroutine omp_destroy_lock(svar) integer (kind=omp_lock_kind) svar subroutine omp_destroy_nest_lock(nvar) integer (kind=omp_nest_lock_kind) nvar Fortran 11 12 Constraints on Arguments 13 14 A program that accesses a lock that is not in the unlocked state through either routine is non-conforming. 144 OpenMP API • Version 3.1 July 2011 1 Effect 2 The effect of these routines is to change the state of the lock to uninitialized. 3 3.3.3 omp_set_lock and omp_set_nest_lock 4 Summary 5 6 These routines provide a means of setting an OpenMP lock. The calling task region is suspended until the lock is set. 7 Format C/C++ void omp_set_lock(omp_lock_t *lock); void omp_set_nest_lock(omp_nest_lock_t *lock); C/C++ 8 Fortran subroutine omp_set_lock(svar) integer (kind=omp_lock_kind) svar subroutine omp_set_nest_lock(nvar) integer (kind=omp_nest_lock_kind) nvar Fortran 9 10 Constraints on Arguments 11 12 13 A program that accesses a lock that is in the uninitialized state through either routine is non-conforming. A simple lock accessed by omp_set_lock that is in the locked state must not be owned by the task that contains the call or deadlock will result. Chapter 3 Runtime Library Routines 145 1 Effect 2 3 Each of these routines causes suspension of the task executing the routine until the specified lock is available and then sets the lock. 4 5 A simple lock is available if it is unlocked. Ownership of the lock is granted to the task executing the routine. 6 7 8 A nestable lock is available if it is unlocked or if it is already owned by the task executing the routine. The task executing the routine is granted, or retains, ownership of the lock, and the nesting count for the lock is incremented. 9 3.3.4 omp_unset_lock and omp_unset_nest_lock 10 Summary 11 These routines provide the means of unsetting an OpenMP lock. 12 Format C/C++ void omp_unset_lock(omp_lock_t *lock); void omp_unset_nest_lock(omp_nest_lock_t *lock); C/C++ 13 Fortran subroutine omp_unset_lock(svar) integer (kind=omp_lock_kind) svar subroutine omp_unset_nest_lock(nvar) integer (kind=omp_nest_lock_kind) nvar Fortran 14 146 OpenMP API • Version 3.1 July 2011 1 Constraints on Arguments 2 3 A program that accesses a lock that is not in the locked state or that is not owned by the task that contains the call through either routine is non-conforming. 4 Effect 5 For a simple lock, the omp_unset_lock routine causes the lock to become unlocked. 6 7 For a nestable lock, the omp_unset_nest_lock routine decrements the nesting count, and causes the lock to become unlocked if the resulting nesting count is zero. For either routine, if the lock becomes unlocked, and if one or more task regions were suspended because the lock was unavailable, the effect is that one task is chosen and given ownership of the lock. 8 9 10 11 3.3.5 omp_test_lock and omp_test_nest_lock 12 Summary 13 14 These routines attempt to set an OpenMP lock but do not suspend execution of the task executing the routine. 15 Format C/C++ int omp_test_lock(omp_lock_t *lock); int omp_test_nest_lock(omp_nest_lock_t *lock); C/C++ 16 Fortran logical integer integer integer 17 function omp_test_lock(svar) (kind=omp_lock_kind) svar function omp_test_nest_lock(nvar) (kind=omp_nest_lock_kind) nvar Fortran Chapter 3 Runtime Library Routines 147 1 Constraints on Arguments 2 3 4 A program that accesses a lock that is in the uninitialized state through either routine is non-conforming. The behavior is unspecified if a simple lock accessed by omp_test_lock is in the locked state and is owned by the task that contains the call. 5 Effect 6 7 8 These routines attempt to set a lock in the same manner as omp_set_lock and omp_set_nest_lock, except that they do not suspend execution of the task executing the routine. 9 10 For a simple lock, the omp_test_lock routine returns true if the lock is successfully set; otherwise, it returns false. 11 12 For a nestable lock, the omp_test_nest_lock routine returns the new nesting count if the lock is successfully set; otherwise, it returns zero. 13 3.4 Timing Routines 14 The routines described in this section support a portable wall clock timer. 15 • the omp_get_wtime routine. 16 • the omp_get_wtick routine. 17 3.4.1 omp_get_wtime 18 Summary 19 The omp_get_wtime routine returns elapsed wall clock time in seconds. 148 OpenMP API • Version 3.1 July 2011 1 Format C/C++ double omp_get_wtime(void); C/C++ 2 Fortran double precision function omp_get_wtime() Fortran 3 4 Binding 5 6 The binding thread set for an omp_get_wtime region is the encountering thread. The routine’s return value is not guaranteed to be consistent across any set of threads. 7 Effect 8 9 10 11 12 The omp_get_wtime routine returns a value equal to the elapsed wall clock time in seconds since some “time in the past”. The actual “time in the past” is arbitrary, but it is guaranteed not to change during the execution of the application program. The time returned is a “per-thread time”, so it is not required to be globally consistent across all the threads participating in an application. 13 14 Note – It is anticipated that the routine will be used to measure elapsed times as shown in the following example: C/C++ double start; double end; start = omp_get_wtime(); ... work to be timed ... end = omp_get_wtime(); printf("Work took %f seconds\n", end - start); 15 C/C++ Chapter 3 Runtime Library Routines 149 Fortran 1 DOUBLE PRECISION START, END START = omp_get_wtime() ... work to be timed ... END = omp_get_wtime() PRINT *, "Work took", END - START, "seconds" Fortran 2 3 3.4.2 omp_get_wtick 4 Summary 5 6 The omp_get_wtick routine returns the precision of the timer used by omp_get_wtime. 7 Format C/C++ double omp_get_wtick(void); C/C++ 8 Fortran double precision function omp_get_wtick() Fortran 9 10 Binding 11 12 The binding thread set for an omp_get_wtick region is the encountering thread. The routine’s return value is not guaranteed to be consistent across any set of threads. 150 OpenMP API • Version 3.1 July 2011 1 Effect 2 3 The omp_get_wtick routine returns a value equal to the number of seconds between successive clock ticks of the timer used by omp_get_wtime. Chapter 3 Runtime Library Routines 151 1 This page intentionally left blank. 2 152 OpenMP API • Version 3.1 July 2011 1 2 CHAPTER 4 Environment Variables 3 4 5 6 7 8 9 10 This chapter describes the OpenMP environment variables that specify the settings of the ICVs that affect the execution of OpenMP programs (see Section 2.3 on page 28). The names of the environment variables must be upper case. The values assigned to the environment variables are case insensitive and may have leading and trailing white space. Modifications to the environment variables after the program has started, even if modified by the program itself, are ignored by the OpenMP implementation. However, the settings of some of the ICVs can be modified during the execution of the OpenMP program by the use of the appropriate directive clauses or OpenMP API routines. 11 The environment variables are as follows: 12 13 • OMP_SCHEDULE sets the run-sched-var ICV that specifies the runtime schedule type 14 15 • OMP_NUM_THREADS sets the nthreads-var ICV that specifies the number of threads 16 17 • OMP_DYNAMIC sets the dyn-var ICV that specifies the dynamic adjustment of 18 19 • OMP_PROC_BIND sets the bind-var ICV that controls whether threads are bound to 20 • OMP_NESTED sets the nest-var ICV that enables or disables nested parallelism. 21 22 • OMP_STACKSIZE sets the stacksize-var ICV that specifies the size of the stack for 23 24 • OMP_WAIT_POLICY sets the wait-policy-var ICV that controls the desired behavior 25 26 • OMP_MAX_ACTIVE_LEVELS sets the max-active-levels-var ICV that controls the 27 28 • OMP_THREAD_LIMIT sets the thread-limit-var ICV that controls the maximum and chunk size. It can be set to any of the valid OpenMP schedule types. to use for parallel regions. threads to use for parallel regions. processors. threads created by the OpenMP implementation. of waiting threads. maximum number of nested active parallel regions. number of threads participating in the OpenMP program. 153 1 2 3 The examples in this chapter only demonstrate how these variables might be set in Unix C shell (csh) environments. In Korn shell (ksh) and DOS environments the actions are similar, as follows: 4 • csh: setenv OMP_SCHEDULE "dynamic" • ksh: 5 export OMP_SCHEDULE="dynamic" • DOS: 6 set OMP_SCHEDULE=dynamic 7 4.1 OMP_SCHEDULE 8 9 10 The OMP_SCHEDULE environment variable controls the schedule type and chunk size of all loop directives that have the schedule type runtime, by setting the value of the run-sched-var ICV. 11 The value of this environment variable takes the form: 12 type[,chunk] 13 where 14 • type is one of static, dynamic, guided, or auto 15 • chunk is an optional positive integer that specifies the chunk size 16 17 If chunk is present, there may be white space on either side of the “,”. See Section 2.5.1 on page 39 for a detailed description of the schedule types. 18 19 The behavior of the program is implementation defined if the value of OMP_SCHEDULE does not conform to the above format. 20 21 22 Implementation specific schedules cannot be specified in OMP_SCHEDULE. They can only be specified by calling omp_set_schedule, described in Section 3.2.11 on page 128. 154 OpenMP API • Version 3.1 July 2011 Example: 1 setenv OMP_SCHEDULE "guided,4" setenv OMP_SCHEDULE "dynamic" 2 Cross References 3 • run-sched-var ICV, see Section 2.3 on page 28. 4 • Loop construct, see Section 2.5.1 on page 39. 5 • Parallel loop construct, see Section 2.6.1 on page 56. 6 • omp_set_schedule routine, see Section 3.2.11 on page 128. 7 • omp_get_schedule routine, see Section 3.2.12 on page 130. 8 4.2 OMP_NUM_THREADS 9 10 11 12 13 14 15 The OMP_NUM_THREADS environment variable sets the number of threads to use for parallel regions by setting the initial value of the nthreads-var ICV. See Section 2.3 on page 28 for a comprehensive set of rules about the interaction between the OMP_NUM_THREADS environment variable, the num_threads clause, the omp_set_num_threads library routine and dynamic adjustment of threads, and Section 2.4.1 on page 36 for a complete algorithm that describes how the number of threads for a parallel region is determined. 16 17 18 The value of this environment variable must be a list of positive integer values. The values of the list set the number of threads to use for parallel regions at the corresponding nested level. 19 20 21 The behavior of the program is implementation defined if any value of the list specified in the OMP_NUM_THREADS environment variable leads to a number of threads which is greater than an implementation can support, or if any value is not a positive integer. 22 Example: setenv OMP_NUM_THREADS 4,3,2 23 Cross References: 24 • nthreads-var ICV, see Section 2.3 on page 28. 25 • num_threads clause, Section 2.4 on page 33. Chapter 4 Environment Variables 155 1 • omp_set_num_threads routine, see Section 3.2.1 on page 116. 2 • omp_get_num_threads routine, see Section 3.2.2 on page 117. 3 • omp_get_max_threads routine, see Section 3.2.3 on page 118. 4 • omp_get_team_size routine, see Section 3.2.18 on page 137. 5 6 4.3 OMP_DYNAMIC 7 8 9 10 11 12 13 14 The OMP_DYNAMIC environment variable controls dynamic adjustment of the number of threads to use for executing parallel regions by setting the initial value of the dyn-var ICV. The value of this environment variable must be true or false. If the environment variable is set to true, the OpenMP implementation may adjust the number of threads to use for executing parallel regions in order to optimize the use of system resources. If the environment variable is set to false, the dynamic adjustment of the number of threads is disabled. The behavior of the program is implementation defined if the value of OMP_DYNAMIC is neither true nor false. 15 Example: setenv OMP_DYNAMIC true 16 Cross References: 17 • dyn-var ICV, see Section 2.3 on page 28. 18 • omp_set_dynamic routine, see Section 3.2.7 on page 123. 19 • omp_get_dynamic routine, see Section 3.2.8 on page 124. 20 4.4 OMP_PROC_BIND The OMP_PROC_BIND environment variable sets the value of the global bind-var ICV. The value of this environment variable must be true or false. If the environment variable is set to true, the execution environment should not move OpenMP threads between processors. If the environment variable is set to false, the execution environment may move OpenMP threads between processors. The behavior of the program is implementation defined if the value of OMP_PROC_BIND is neither true nor false. 21 22 23 24 25 26 27 156 OpenMP API • Version 3.1 July 2011 Example: 1 setenv OMP_PROC_BIND true 2 Cross References: 3 • bind-var ICV, see Section 2.3 on page 28. 4 4.5 OMP_NESTED The OMP_NESTED environment variable controls nested parallelism by setting the initial value of the nest-var ICV. The value of this environment variable must be true or false. If the environment variable is set to true, nested parallelism is enabled; if set to false, nested parallelism is disabled. The behavior of the program is implementation defined if the value of OMP_NESTED is neither true nor false. 5 6 7 8 9 Example: 10 setenv OMP_NESTED false 11 Cross References 12 • nest-var ICV, see Section 2.3 on page 28. 13 • omp_set_nested routine, see Section 3.2.9 on page 125. 14 • omp_get_nested routine, see Section 3.2.18 on page 137. 15 16 4.6 OMP_STACKSIZE 17 18 19 The OMP_STACKSIZE environment variable controls the size of the stack for threads created by the OpenMP implementation, by setting the value of the stacksize-var ICV. The environment variable does not control the size of the stack for the initial thread. 20 The value of this environment variable takes the form: 21 size | sizeB | sizeK | sizeM | sizeG 22 where: Chapter 4 Environment Variables 157 1 2 • size is a positive integer that specifies the size of the stack for threads that are created 3 4 5 6 • B, K, M, and G are letters that specify whether the given size is in Bytes, Kilobytes 7 8 If only size is specified and none of B, K, M, or G is specified, then size is assumed to be in Kilobytes. 9 10 11 The behavior of the program is implementation defined if OMP_STACKSIZE does not conform to the above format, or if the implementation cannot provide a stack with the requested size. 12 Examples: by the OpenMP implementation. (1024 Bytes), Megabytes (1024 Kilobytes), or Gigabytes (1024 Megabytes), respectively. If one of these letters is present, there may be white space between size and the letter. setenv setenv setenv setenv setenv setenv setenv OMP_STACKSIZE OMP_STACKSIZE OMP_STACKSIZE OMP_STACKSIZE OMP_STACKSIZE OMP_STACKSIZE OMP_STACKSIZE 2000500B "3000 k " 10M " 10 M " "20 m " " 1G" 20000 13 Cross References 14 • stacksize-var ICV, see Section 2.3 on page 28. 15 4.7 OMP_WAIT_POLICY 16 17 18 19 The OMP_WAIT_POLICY environment variable provides a hint to an OpenMP implementation about the desired behavior of waiting threads by setting the wait-policyvar ICV. A compliant OpenMP implementation may or may not abide by the setting of the environment variable. 20 The value of this environment variable takes the form: 21 ACTIVE | PASSIVE 22 23 24 The ACTIVE value specifies that waiting threads should mostly be active, consuming processor cycles, while waiting. An OpenMP implementation may, for example, make waiting threads spin. 158 OpenMP API • Version 3.1 July 2011 1 2 3 The PASSIVE value specifies that waiting threads should mostly be passive, not consuming processor cycles, while waiting. For example, an OpenMP implementation may make waiting threads yield the processor to other threads or go to sleep. 4 The details of the ACTIVE and PASSIVE behaviors are implementation defined. 5 Examples: setenv setenv setenv setenv OMP_WAIT_POLICY OMP_WAIT_POLICY OMP_WAIT_POLICY OMP_WAIT_POLICY ACTIVE active PASSIVE passive 6 Cross References 7 • wait-policy-var ICV, see Section 2.3 on page 24. 8 4.8 OMP_MAX_ACTIVE_LEVELS 9 10 11 The OMP_MAX_ACTIVE_LEVELS environment variable controls the maximum number of nested active parallel regions by setting the initial value of the max-active-levels-var ICV. 12 13 14 15 16 The value of this environment variable must be a non-negative integer. The behavior of the program is implementation defined if the requested value of OMP_MAX_ACTIVE_LEVELS is greater than the maximum number of nested active parallel levels an implementation can support, or if the value is not a non-negative integer. 17 Cross References 18 • max-active-levels-var ICV, see Section 2.3 on page 28. 19 • omp_set_max_active_levels routine, see Section 3.2.14 on page 132. 20 • omp_get_max_active_levels routine, see Section 3.2.15 on page 134. Chapter 4 Environment Variables 159 1 4.9 OMP_THREAD_LIMIT 2 3 The OMP_THREAD_LIMIT environment variable sets the number of OpenMP threads to use for the whole OpenMP program by setting the thread-limit-var ICV. 4 5 6 7 The value of this environment variable must be a positive integer. The behavior of the program is implementation defined if the requested value of OMP_THREAD_LIMIT is greater than the number of threads an implementation can support, or if the value is not a positive integer. 8 Cross References 9 • thread-limit-var ICV, see Section 2.3 on page 28. • omp_get_thread_limit routine 10 160 OpenMP API • Version 3.1 July 2011 1 APPENDIX A 2 Examples 3 The following are examples of the constructs and routines defined in this document. 4 5 A statement following a directive is compound only when necessary, and a noncompound statement is indented with respect to a directive preceding it. C/C++ C/C++ 6 7 8 9 A.1 A Simple Parallel Loop The following example demonstrates how to parallelize a simple loop using the parallel loop construct (Section 2.6.1 on page 56). The loop iteration variable is private by default, so it is not necessary to specify it explicitly in a private clause. C/C++ 10 Example A.1.1c 11 12 13 14 15 16 17 18 void simple(int n, float *a, float *b) { int i; #pragma omp parallel for for (i=1; i<n; i++) /* i is private by default */ b[i] = (a[i] + a[i-1]) / 2.0; } C/C++ 161 Fortran Example A.1.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 SUBROUTINE SIMPLE(N, A, B) INTEGER I, N REAL B(N), A(N) !$OMP PARALLEL DO !I is private by default DO I=2,N B(I) = (A(I) + A(I-1)) / 2.0 ENDDO !$OMP END PARALLEL DO END SUBROUTINE SIMPLE Fortran 14 A.2 The OpenMP Memory Model In the following example, at Print 1, the value of x could be either 2 or 5, depending on the timing of the threads, and the implementation of the assignment to x. There are two reasons that the value at Print 1 might not be 5. First, Print 1 might be executed before the assignment to x is executed. Second, even if Print 1 is executed after the assignment, the value 5 is not guaranteed to be seen by thread 1 because a flush may not have been executed by thread 0 since the assignment. 15 16 17 18 19 20 162 OpenMP API • Version 3.1 July 2011 1 2 3 The barrier after Print 1 contains implicit flushes on all threads, as well as a thread synchronization, so the programmer is guaranteed that the value 5 will be printed by both Print 2 and Print 3. 4 Example A.2.1c 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 C/C++ #include <stdio.h> #include <omp.h> int main(){ int x; x = 2; #pragma omp parallel num_threads(2) shared(x) { if (omp_get_thread_num() == 0) { x = 5; } else { /* Print 1: the following read of x has a race */ printf("1: Thread# %d: x = %d\n", omp_get_thread_num(),x ); } #pragma omp barrier if (omp_get_thread_num() == 0) { /* Print 2 */ printf("2: Thread# %d: x = %d\n", omp_get_thread_num(),x ); } else { /* Print 3 */ printf("3: Thread# %d: x = %d\n", omp_get_thread_num(),x ); } } return 0; } C/C++ Appendix A Examples 163 Fortran Example A.2.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 PROGRAM MEMMODEL INCLUDE "omp_lib.h" INTEGER X ! or USE OMP_LIB X = 2 !$OMP PARALLEL NUM_THREADS(2) SHARED(X) IF (OMP_GET_THREAD_NUM() .EQ. 0) THEN X = 5 ELSE ! PRINT 1: The following read of x has a race PRINT *,"1: THREAD# ", OMP_GET_THREAD_NUM(), "X = ", X ENDIF !$OMP BARRIER IF (OMP_GET_THREAD_NUM() .EQ. 0) THEN ! PRINT 2 PRINT *,"2: THREAD# ", OMP_GET_THREAD_NUM(), "X = ", X ELSE ! PRINT 3 PRINT *,"3: THREAD# ", OMP_GET_THREAD_NUM(), "X = ", X ENDIF !$OMP END PARALLEL END PROGRAM MEMMODEL Fortran The following example demonstrates why synchronization is difficult to perform correctly through variables. The value of flag is undefined in both prints on thread 1 and the value of data is only well-defined in the second print. 29 30 31 164 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 Example A.2.2c C/C++ #include <omp.h> #include <stdio.h> int main() { int data; int flag=0; #pragma omp parallel num_threads(2) { if (omp_get_thread_num()==0) { /* Write to the data buffer that will be read by thread */ data = 42; /* Flush data to thread 1 and strictly order the write to data relative to the write to the flag */ #pragma omp flush(flag, data) /* Set flag to release thread 1 */ flag = 1; /* Flush flag to ensure that thread 1 sees the change */ #pragma omp flush(flag) } else if(omp_get_thread_num()==1) { /* Loop until we see the update to the flag */ #pragma omp flush(flag, data) while (flag < 1) { #pragma omp flush(flag, data) } /* Values of flag and data are undefined */ printf("flag=%d data=%d\n", flag, data); #pragma omp flush(flag, data) /* Values data will be 42, value of flag still undefined */ printf("flag=%d data=%d\n", flag, data); } } return 0; } C/C++ 43 Appendix A Examples 165 Fortran Example A.2.2f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 PROGRAM INCLUDE INTEGER INTEGER EXAMPLE "omp_lib.h" ! or USE OMP_LIB DATA FLAG FLAG = 0 !$OMP PARALLEL NUM_THREADS(2) IF(OMP_GET_THREAD_NUM() .EQ. 0) THEN ! Write to the data buffer that will be read by thread 1 DATA = 42 ! Flush DATA to thread 1 and strictly order the write to DATA ! relative to the write to the FLAG !$OMP FLUSH(FLAG, DATA) ! Set FLAG to release thread 1 FLAG = 1; ! Flush FLAG to ensure that thread 1 sees the change */ !$OMP FLUSH(FLAG) ELSE IF(OMP_GET_THREAD_NUM() .EQ. 1) THEN ! Loop until we see the update to the FLAG !$OMP FLUSH(FLAG, DATA) DO WHILE(FLAG .LT. 1) !$OMP FLUSH(FLAG, DATA) ENDDO ! Values of FLAG and DATA are undefined PRINT *, 'FLAG=', FLAG, ' DATA=', DATA !$OMP FLUSH(FLAG, DATA) !Values DATA will be 42, value of FLAG still undefined */ PRINT *, 'FLAG=', FLAG, ' DATA=', DATA ENDIF !$OMP END PARALLEL END Fortran The next example demonstrates why synchronization is difficult to perform correctly through variables. Because the write(1)-flush(1)-flush(2)-read(2) sequence cannot be guaranteed in the example, the statements on thread 0 and thread 1 may execute in either order. 35 36 37 38 166 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 C/C++ Example A.2.3c #include <omp.h> #include <stdio.h> int main() { int flag=0; #pragma omp parallel num_threads(3) { if(omp_get_thread_num()==0) { /* Set flag to release thread 1 */ #pragma omp atomic update flag++; /* Flush of flag is implied by the atomic directive */ } else if(omp_get_thread_num()==1) { /* Loop until we see that flag reaches 1*/ #pragma omp flush(flag) while(flag < 1) { #pragma omp flush(flag) } printf("Thread 1 awoken\n"); /* Set flag to release thread 2 */ #pragma omp atomic update flag++; /* Flush of flag is implied by the atomic directive */ } else if(omp_get_thread_num()==2) { /* Loop until we see that flag reaches 2 */ #pragma omp flush(flag) while(flag < 2) { #pragma omp flush(flag) } printf("Thread 2 awoken\n"); } } return 0; } C/C++ 45 Appendix A Examples 167 Fortran Example A.2.3f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 PROGRAM EXAMPLE INCLUDE "omp_lib.h" ! or USE OMP_LIB INTEGER FLAG FLAG = 0 !$OMP PARALLEL NUM_THREADS(3) IF(OMP_GET_THREAD_NUM() .EQ. 0) THEN ! Set flag to release thread 1 !$OMP ATOMIC UPDATE FLAG = FLAG + 1 !Flush of FLAG is implied by the atomic directive ELSE IF(OMP_GET_THREAD_NUM() .EQ. 1) THEN ! Loop until we see that FLAG reaches 1 !$OMP FLUSH(FLAG, DATA) DO WHILE(FLAG .LT. 1) !$OMP FLUSH(FLAG, DATA) ENDDO PRINT *, 'Thread 1 awoken' ! Set FLAG to release thread 2 !$OMP ATOMIC UPDATE FLAG = FLAG + 1 !Flush of FLAG is implied by the atomic directive ELSE IF(OMP_GET_THREAD_NUM() .EQ. 2) THEN ! Loop until we see that FLAG reaches 2 !$OMP FLUSH(FLAG, DATA) DO WHILE(FLAG .LT. 2) !$OMP FLUSH(FLAG, DATA) ENDDO PRINT *, 'Thread 2 awoken' ENDIF !$OMP END PARALLEL END Fortran 37 168 OpenMP API • Version 3.1 July 2011 1 A.3 Conditional Compilation C/C++ 2 3 4 The following example illustrates the use of conditional compilation using the OpenMP macro _OPENMP (Section 2.2 on page 26). With OpenMP compilation, the _OPENMP macro becomes defined. 5 Example A.3.1c 6 7 8 9 10 11 12 13 14 15 16 #include <stdio.h> int main() { # ifdef _OPENMP printf("Compiled by an OpenMP-compliant implementation.\n"); # endif return 0; } C/C++ Fortran 17 18 19 20 The following example illustrates the use of the conditional compilation sentinel (see Section 2.2 on page 26). With OpenMP compilation, the conditional compilation sentinel !$ is recognized and treated as two spaces. In fixed form source, statements guarded by the sentinel must start after column 6. 21 Example A.3.1f 22 23 24 25 26 27 PROGRAM EXAMPLE C234567890 !$ PRINT *, "Compiled by an OpenMP-compliant implementation." END PROGRAM EXAMPLE Fortran Appendix A Examples 169 1 A.4 Internal Control Variables (ICVs) 2 3 4 5 6 7 According to Section 2.3 on page 28, an OpenMP implementation must act as if there are ICVs that control the behavior of the program. This example illustrates two ICVs, nthreads-var and max-active-levels-var. The nthreads-var ICV controls the number of threads requested for encountered parallel regions; there is one copy of this ICV per task. The max-active-levels-var ICV controls the maximum number of nested active parallel regions; there is one copy of this ICV for the whole program. 8 9 10 11 12 13 In the following example, the nest-var, max-active-levels-var, dyn-var, and nthreads-var ICVs are modified through calls to the runtime library routines omp_set_nested, omp_set_max_active_levels, omp_set_dynamic, and omp_set_num_threads respectively. These ICVs affect the operation of parallel regions. Each implicit task generated by a parallel region has its own copy of the nest-var, dyn-var, and nthreads-var ICVs. 14 15 16 17 In the following example, the new value of nthreads-var applies only to the implicit tasks that execute the call to omp_set_num_threads. There is one copy of the maxactive-levels-var ICV for the whole program and its value is the same for all tasks. This example assumes that nested parallelism is supported. 18 19 The outer parallel region creates a team of two threads; each of the threads will execute one of the two implicit tasks generated by the outer parallel region. 20 21 22 23 24 Each implicit task generated by the outer parallel region calls omp_set_num_threads(3), assigning the value 3 to its respective copy of nthreads-var. Then each implicit task encounters an inner parallel region that creates a team of three threads; each of the threads will execute one of the three implicit tasks generated by that inner parallel region. 25 26 Since the outer parallel region is executed by 2 threads, and the inner by 3, there will be a total of 6 implicit tasks generated by the two inner parallel regions. 27 28 29 Each implicit task generated by an inner parallel region will execute the call to omp_set_num_threads(4), assigning the value 4 to its respective copy of nthreads-var. 30 31 The print statement in the outer parallel region is executed by only one of the threads in the team. So it will be executed only once. 32 33 34 The print statement in an inner parallel region is also executed by only one of the threads in the team. Since we have a total of two inner parallel regions, the print statement will be executed twice -- once per inner parallel region. 170 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 Example A.4.1c C/C++ #include <stdio.h> #include <omp.h> int main (void) { omp_set_nested(1); omp_set_max_active_levels(8); omp_set_dynamic(0); omp_set_num_threads(2); #pragma omp parallel { omp_set_num_threads(3); #pragma omp parallel { omp_set_num_threads(4); #pragma omp single { /* * The following should print: * Inner: max_act_lev=8, num_thds=3, max_thds=4 * Inner: max_act_lev=8, num_thds=3, max_thds=4 */ printf ("Inner: max_act_lev=%d, num_thds=%d, max_thds=%d\n", omp_get_max_active_levels(), omp_get_num_threads(), omp_get_max_threads()); } } #pragma omp barrier #pragma omp single { /* * The following should print: * Outer: max_act_lev=8, num_thds=2, max_thds=3 */ printf ("Outer: max_act_lev=%d, num_thds=%d, max_thds=%d\n", omp_get_max_active_levels(), omp_get_num_threads(), omp_get_max_threads()); } } return 0; } C/C++ Appendix A Examples 171 Fortran Example A.4.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 program icv use omp_lib call call call call omp_set_nested(.true.) omp_set_max_active_levels(8) omp_set_dynamic(.false.) omp_set_num_threads(2) !$omp parallel call omp_set_num_threads(3) !$omp parallel call omp_set_num_threads(4) !$omp single ! The following should print: ! Inner: max_act_lev= 8 , num_thds= 3 , max_thds= 4 ! Inner: max_act_lev= 8 , num_thds= 3 , max_thds= 4 print *, "Inner: max_act_lev=", omp_get_max_active_levels(), & ", num_thds=", omp_get_num_threads(), & ", max_thds=", omp_get_max_threads() !$omp end single !$omp end parallel !$omp barrier !$omp single ! The following should print: ! Outer: max_act_lev= 8 , num_thds= 2 , max_thds= 3 print *, "Outer: max_act_lev=", omp_get_max_active_levels(), & ", num_thds=", omp_get_num_threads(), & ", max_thds=", omp_get_max_threads() !$omp end single !$omp end parallel end Fortran 35 A.5 The parallel Construct The parallel construct (Section 2.4 on page 33) can be used in coarse-grain parallel programs. In the following example, each thread in the parallel region decides what part of the global array x to work on, based on the thread number: 36 37 38 172 OpenMP API • Version 3.1 July 2011 1 Example A.5.1c 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 #include <omp.h> C/C++ void subdomain(float *x, int istart, int ipoints) { int i; for (i = 0; i < ipoints; i++) x[istart+i] = 123.456; } void sub(float *x, int npoints) { int iam, nt, ipoints, istart; #pragma omp parallel default(shared) private(iam,nt,ipoints,istart) { iam = omp_get_thread_num(); nt = omp_get_num_threads(); ipoints = npoints / nt; /* size of partition */ istart = iam * ipoints; /* starting array index */ if (iam == nt-1) /* last thread may do more */ ipoints = npoints - istart; subdomain(x, istart, ipoints); } } int main() { float array[10000]; sub(array, 10000); return 0; } C/C++ Appendix A Examples 173 Fortran Example A.5.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 SUBROUTINE SUBDOMAIN(X, ISTART, IPOINTS) INTEGER ISTART, IPOINTS REAL X(*) INTEGER I 100 DO 100 I=1,IPOINTS X(ISTART+I) = 123.456 CONTINUE END SUBROUTINE SUBDOMAIN SUBROUTINE SUB(X, NPOINTS) INCLUDE "omp_lib.h" ! or USE OMP_LIB REAL X(*) INTEGER NPOINTS INTEGER IAM, NT, IPOINTS, ISTART !$OMP PARALLEL DEFAULT(PRIVATE) SHARED(X,NPOINTS) IAM = OMP_GET_THREAD_NUM() NT = OMP_GET_NUM_THREADS() IPOINTS = NPOINTS/NT ISTART = IAM * IPOINTS IF (IAM .EQ. NT-1) THEN IPOINTS = NPOINTS - ISTART ENDIF CALL SUBDOMAIN(X,ISTART,IPOINTS) !$OMP END PARALLEL END SUBROUTINE SUB PROGRAM PAREXAMPLE REAL ARRAY(10000) CALL SUB(ARRAY, 10000) END PROGRAM PAREXAMPLE Fortran 174 OpenMP API • Version 3.1 July 2011 2 Controlling the Number of Threads on Multiple Nesting Levels 3 4 5 The following examples demonstrate how to use the OMP_NUM_THREADS environment variable (Section 2.3.2 on page 29) to control the number of threads on multiple nesting levels: 6 Example A.6.1c 1 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 A.6 C/C++ #include <stdio.h> #include <omp.h> int main (void) { omp_set_nested(1); omp_set_dynamic(0); #pragma omp parallel { #pragma omp parallel { #pragma omp single { /* * If OMP_NUM_THREADS=2,3 was set, the following should print: * Inner: num_thds=3 * Inner: num_thds=3 * * If nesting is not supported, the following should print: * Inner: num_thds=1 * Inner: num_thds=1 */ printf ("Inner: num_thds=%d\n", omp_get_num_threads()); } } #pragma omp barrier omp_set_nested(0); #pragma omp parallel { #pragma omp single { /* * Even if OMP_NUM_THREADS=2,3 was set, the following should * print, because nesting is disabled: * Inner: num_thds=1 * Inner: num_thds=1 */ printf ("Inner: num_thds=%d\n", omp_get_num_threads()); Appendix A Examples 175 1 2 3 4 5 6 7 8 9 10 11 12 13 14 } } #pragma omp barrier #pragma omp single { /* * If OMP_NUM_THREADS=2,3 was set, the following should print: * Outer: num_thds=2 */ printf ("Outer: num_thds=%d\n", omp_get_num_threads()); } } return 0; } C/C++ Fortran Example A.6.1f 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 !$omp !$omp !$omp !$omp !$omp !$omp !$omp !$omp !$omp !$omp !$omp !$omp 176 program icv use omp_lib call omp_set_nested(.true.) call omp_set_dynamic(.false.) parallel parallel single ! If OMP_NUM_THREADS=2,3 was set, the following should print: ! Inner: num_thds= 3 ! Inner: num_thds= 3 ! If nesting is not supported, the following should print: ! Inner: num_thds= 1 ! Inner: num_thds= 1 print *, "Inner: num_thds=", omp_get_num_threads() end single end parallel barrier call omp_set_nested(.false.) parallel single ! Even if OMP_NUM_THREADS=2,3 was set, the following should print, ! because nesting is disabled: ! Inner: num_thds= 1 ! Inner: num_thds= 1 print *, "Inner: num_thds=", omp_get_num_threads() end single end parallel barrier single ! If OMP_NUM_THREADS=2,3 was set, the following should print: ! Outer: num_thds= 2 print *, "Outer: num_thds=", omp_get_num_threads() OpenMP API • Version 3.1 July 2011 1 2 3 !$omp end single !$omp end parallel end Fortran 4 5 6 7 A.7 Interaction Between the num_threads Clause and omp_set_dynamic 8 9 10 The following example demonstrates the num_threads clause (Section 2.4 on page 33) and the effect of the omp_set_dynamic routine (Section 3.2.7 on page 123) on it. 11 12 13 14 15 The call to the omp_set_dynamic routine with argument 0 in C/C++, or .FALSE. in Fortran, disables the dynamic adjustment of the number of threads in OpenMP implementations that support it. In this case, 10 threads are provided. Note that in case of an error the OpenMP implementation is free to abort the program or to supply any number of threads available. 16 C/C++ 17 Example A.7.1c 18 19 20 21 22 23 24 25 26 27 #include <omp.h> int main() { omp_set_dynamic(0); #pragma omp parallel num_threads(10) { /* do work here */ } return 0; } C/C++ Fortran 28 29 30 Example A.7.1f PROGRAM EXAMPLE INCLUDE "omp_lib.h" ! or USE OMP_LIB Appendix A Examples 177 1 2 3 4 5 CALL OMP_SET_DYNAMIC(.FALSE.) PARALLEL NUM_THREADS(10) ! do work here !$OMP END PARALLEL END PROGRAM EXAMPLE !$OMP Fortran 6 7 8 The call to the omp_set_dynamic routine with a non-zero argument in C/C++, or .TRUE. in Fortran, allows the OpenMP implementation to choose any number of threads between 1 and 10 (see also Algorithm 2.1 in Section 2.4.1 on page 36). 9 Example A.7.2c 10 11 12 13 14 15 16 17 18 19 C/C++ #include <omp.h> int main() { omp_set_dynamic(1); #pragma omp parallel num_threads(10) { /* do work here */ } return 0; } C/C++ Fortran 20 Example A.7.2f 21 22 23 24 25 26 27 PROGRAM EXAMPLE INCLUDE "omp_lib.h" ! or USE OMP_LIB CALL OMP_SET_DYNAMIC(.TRUE.) !$OMP PARALLEL NUM_THREADS(10) ! do work here !$OMP END PARALLEL END PROGRAM EXAMPLE Fortran It is good practice to set the dyn-var ICV explicitly by calling the omp_set_dynamic routine, as its default setting is implementation defined. 28 29 178 OpenMP API • Version 3.1 July 2011 1 Fortran 2 3 4 5 6 A.8 Fortran Restrictions on the do Construct If an end do directive follows a do-construct in which several DO statements share a DO termination statement, then a do directive can only be specified for the outermost of these DO statements. For more information, see Section 2.5.1 on page 39. The following example contains correct usages of loop constructs: Appendix A Examples 179 Example A.8.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 SUBROUTINE WORK(I, J) INTEGER I,J END SUBROUTINE WORK SUBROUTINE DO_GOOD() INTEGER I, J REAL A(1000) !$OMP 100 !$OMP 200 !$OMP DO 100 I = 1,10 DO DO 100 J = 1,10 CALL WORK(I,J) CONTINUE ! !$OMP ENDDO implied here DO DO 200 J = 1,10 A(I) = I + 1 ENDDO !$OMP DO DO 300 I = 1,10 DO 300 J = 1,10 CALL WORK(I,J) 300 CONTINUE !$OMP ENDDO END SUBROUTINE DO_GOOD 29 30 The following example is non-conforming because the matching do directive for the end do does not precede the outermost loop: 31 Example A.8.2f 32 33 34 35 36 37 38 39 40 41 42 43 44 45 SUBROUTINE WORK(I, J) INTEGER I,J END SUBROUTINE WORK SUBROUTINE DO_WRONG INTEGER I, J DO 100 I = 1,10 DO DO 100 J = 1,10 CALL WORK(I,J) 100 CONTINUE !$OMP ENDDO END SUBROUTINE DO_WRONG !$OMP Fortran 180 OpenMP API • Version 3.1 July 2011 1 Fortran 2 3 A.9 Fortran Private Loop Iteration Variables 4 5 6 7 In general loop iteration variables will be private, when used in the do-loop of a do and parallel do construct or in sequential loops in a parallel construct (see Section 2.5.1 on page 39 and Section 2.9.1 on page 84). In the following example of a sequential loop in a parallel construct the loop iteration variable I will be private. 8 Example A.9.1f 9 10 11 12 13 14 15 16 17 18 19 20 21 22 SUBROUTINE PLOOP_1(A,N) INCLUDE "omp_lib.h" ! or USE OMP_LIB REAL A(*) INTEGER I, MYOFFSET, N !$OMP PARALLEL PRIVATE(MYOFFSET) MYOFFSET = OMP_GET_THREAD_NUM()*N DO I = 1, N A(MYOFFSET+I) = FLOAT(I) ENDDO !$OMP END PARALLEL END SUBROUTINE PLOOP_1 Appendix A Examples 181 1 2 In exceptional cases, loop iteration variables can be made shared, as in the following example: 3 Example A.9.2f 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 SUBROUTINE PLOOP_2(A,B,N,I1,I2) REAL A(*), B(*) INTEGER I1, I2, N 26 27 Note however that the use of shared loop iteration variables can easily lead to race conditions. !$OMP PARALLEL SHARED(A,B,I1,I2) !$OMP SECTIONS !$OMP SECTION DO I1 = I1, N IF (A(I1).NE.0.0) EXIT ENDDO !$OMP SECTION DO I2 = I2, N IF (B(I2).NE.0.0) EXIT ENDDO !$OMP END SECTIONS !$OMP SINGLE IF (I1.LE.N) PRINT *, 'ITEMS IN A UP TO ', I1, 'ARE ALL ZERO.' IF (I2.LE.N) PRINT *, 'ITEMS IN B UP TO ', I2, 'ARE ALL ZERO.' !$OMP END SINGLE !$OMP END PARALLEL END SUBROUTINE PLOOP_2 Fortran 28 A.10 The nowait clause If there are multiple independent loops within a parallel region, you can use the nowait clause (see Section 2.5.1 on page 39) to avoid the implied barrier at the end of the loop construct, as follows: 29 30 31 182 OpenMP API • Version 3.1 July 2011 1 Example A.10.1c 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 #include <math.h> C/C++ void nowait_example(int n, int m, float *a, float *b, float *y, float *z) { int i; #pragma omp parallel { #pragma omp for nowait for (i=1; i<n; i++) b[i] = (a[i] + a[i-1]) / 2.0; #pragma omp for nowait for (i=0; i<m; i++) y[i] = sqrt(z[i]); } } C/C++ Fortran 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 Example A.10.1f SUBROUTINE NOWAIT_EXAMPLE(N, M, A, B, Y, Z) INTEGER N, M REAL A(*), B(*), Y(*), Z(*) INTEGER I !$OMP PARALLEL !$OMP DO DO I=2,N B(I) = (A(I) + A(I-1)) / 2.0 ENDDO !$OMP END DO NOWAIT !$OMP DO DO I=1,M Y(I) = SQRT(Z(I)) ENDDO !$OMP END DO NOWAIT !$OMP END PARALLEL END SUBROUTINE NOWAIT_EXAMPLE Fortran Appendix A Examples 183 1 2 3 4 5 In the following example, static scheduling distributes the same logical iteration numbers to the threads that execute the three loop regions. This allows the nowait clause to be used, even though there is a data dependence between the loops. The dependence is satisfied as long the same thread executes the same logical iteration numbers in each loop. 6 7 8 9 Note that the iteration count of the loops must be the same. The example satisfies this requirement, since the iteration space of the first two loops is from 0 to n-1 (from 1 to N in the Fortran version), while the iteration space of the last loop is from 1 to n (2 to N+1 in the Fortran version). Example A.10.2c 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 C/C++ #include <math.h> void nowait_example2(int n, float *a, float *b, float *c, float *y, float *z) { int i; #pragma omp parallel { #pragma omp for schedule(static) nowait for (i=0; i<n; i++) c[i] = (a[i] + b[i]) / 2.0f; #pragma omp for schedule(static) nowait for (i=0; i<n; i++) z[i] = sqrtf(c[i]); #pragma omp for schedule(static) nowait for (i=1; i<=n; i++) y[i] = z[i-1] + a[i]; } } C/C++ Fortran Example A.10.2f 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 SUBROUTINE NOWAIT_EXAMPLE2(N, A, B, C, Y, Z) INTEGER N REAL A(*), B(*), C(*), Y(*), Z(*) INTEGER I !$OMP PARALLEL !$OMP DO SCHEDULE(STATIC) DO I=1,N C(I) = (A(I) + B(I)) / 2.0 ENDDO !$OMP END DO NOWAIT !$OMP DO SCHEDULE(STATIC) DO I=1,N Z(I) = SQRT(C(I)) ENDDO !$OMP END DO NOWAIT 184 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 !$OMP DO SCHEDULE(STATIC) DO I=2,N+1 Y(I) = Z(I-1) + A(I) ENDDO !$OMP END DO NOWAIT !$OMP END PARALLEL END SUBROUTINE NOWAIT_EXAMPLE2 Fortran 8 A.11 The collapse clause 9 10 11 12 For the following three examples, see Section 2.5.1 on page 39 for a description of the collapse clause, Section 2.8.7 on page 82 for a description of the ordered construct, and Section 2.9.3.5 on page 101 for a description of the lastprivate clause. 13 14 15 16 17 In the following example, the k and j loops are associated with the loop construct. So the iterations of the k and j loops are collapsed into one loop with a larger iteration space, and that loop is then divided among the threads in the current team. Since the i loop is not associated with the loop construct, it is not collapsed, and the i loop is executed sequentially in its entirety in every iteration of the collapsed k and j loop. 18 19 20 21 The variable j can be omitted from the private clause when the collapse clause is used since it is implicitly private. However, if the collapse clause is omitted then j will be shared if it is omitted from the private clause. In either case, k is implicitly private and could be omitted from the private clause. 22 Example A.11.1c 23 24 25 26 27 28 29 30 31 32 33 void bar(float *a, int i, int j, int k); int kl, ku, ks, jl, ju, js, il, iu,is; void sub(float *a) { int i, j, k; #pragma omp for collapse(2) private(i, k, j) for (k=kl; k<=ku; k+=ks) for (j=jl; j<=ju; j+=js) for (i=il; i<=iu; i+=is) bar(a,i,j,k); } C/C++ C/C++ Appendix A Examples 185 Fortran Example A.11.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 subroutine sub(a) real a(*) integer kl, ku, ks, jl, ju, js, il, iu, is common /csub/ kl, ku, ks, jl, ju, js, il, iu, is integer i, j, k !$omp do collapse(2) private(i,j,k) do k = kl, ku, ks do j = jl, ju, js do i = il, iu, is call bar(a,i,j,k) enddo enddo enddo !$omp end do end subroutine Fortran 17 18 19 In the next example, the k and j loops are associated with the loop construct. So the iterations of the k and j loops are collapsed into one loop with a larger iteration space, and that loop is then divided among the threads in the current team. 20 21 22 23 24 The sequential execution of the iterations in the k and j loops determines the order of the iterations in the collapsed iteration space. This implies that in the sequentially last iteration of the collapsed iteration space, k will have the value 2 and j will have the value 3. Since klast and jlast are lastprivate, their values are assigned by the sequentially last iteration of the collapsed k and j loop. This example prints: 2 3. 186 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Example A.11.2c C/C++ #include <stdio.h> void test() { int j, k, jlast, klast; #pragma omp parallel { #pragma omp for collapse(2) lastprivate(jlast, klast) for (k=1; k<=2; k++) for (j=1; j<=3; j++) { jlast=j; klast=k; } #pragma omp single printf("%d %d\n", klast, jlast); } } C/C++ Fortran 19 Example A.11.2f 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 program test !$omp parallel !$omp do private(j,k) collapse(2) lastprivate(jlast, klast) do k = 1,2 do j = 1,3 jlast=j klast=k enddo enddo !$omp end do !$omp single print *, klast, jlast !$omp end single !$omp end parallel end program test Fortran 35 The next example illustrates the interaction of the collapse and ordered clauses. Appendix A Examples 187 1 2 3 4 5 In the example, the loop construct has both a collapse clause and an ordered clause. The collapse clause causes the iterations of the k and j loops to be collapsed into one loop with a larger iteration space, and that loop is divided among the threads in the current team. An ordered clause is added to the loop construct, because an ordered region binds to the loop region arising from the loop construct. 6 7 8 9 10 11 12 According to Section 2.8.7 on page 82, a thread must not execute more than one ordered region that binds to the same loop region. So the collapse clause is required for the example to be conforming. With the collapse clause, the iterations of the k and j loops are collapsed into one loop, and therefore only one ordered region will bind to the collapsed k and j loop. Without the collapse clause, there would be two ordered regions that bind to each iteration of the k loop (one arising from the first iteration of the j loop, and the other arising from the second iteration of the j loop). 13 14 15 16 17 18 19 20 0 0 0 1 1 1 21 Example A.11.3c 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 #include <omp.h> #include <stdio.h> void work(int a, int j, int k); void sub() { int j, k, a; #pragma omp parallel num_threads(2) { #pragma omp for collapse(2) ordered private(j,k) schedule(static,3) for (k=1; k<=3; k++) for (j=1; j<=2; j++) { #pragma omp ordered printf("%d %d %d\n", omp_get_thread_num(), k, j); /* end ordered */ work(a,j,k); } } } The code prints 1 1 2 2 3 3 1 2 1 2 1 2 C/C++ C/C++ 188 OpenMP API • Version 3.1 July 2011 Fortran Example A.11.3f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 !$omp !$omp !$omp !$omp !$omp !$omp program test include 'omp_lib.h' parallel num_threads(2) do collapse(2) ordered private(j,k) schedule(static,3) do k = 1,3 do j = 1,2 ordered print *, omp_get_thread_num(), k, j end ordered call work(a,j,k) enddo enddo end do end parallel end program test Fortran 17 18 A.12 The parallel sections Construct 19 20 21 In the following example (for Section 2.6.2 on page 57) routines XAXIS, YAXIS, and ZAXIS can be executed concurrently. The first section directive is optional. Note that all section directives need to appear in the parallel sections construct. 22 Example A.12.1c 23 24 25 26 27 28 29 30 31 32 33 34 35 36 void XAXIS(); void YAXIS(); void ZAXIS(); C/C++ void sect_example() { #pragma omp parallel sections { #pragma omp section XAXIS(); #pragma omp section YAXIS(); Appendix A Examples 189 1 2 3 4 #pragma omp section ZAXIS(); } } C/C++ Fortran Example A.12.1f 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 SUBROUTINE SECT_EXAMPLE() !$OMP PARALLEL SECTIONS !$OMP SECTION CALL XAXIS() !$OMP SECTION CALL YAXIS() !$OMP SECTION CALL ZAXIS() !$OMP END PARALLEL SECTIONS END SUBROUTINE SECT_EXAMPLE Fortran 22 The firstprivate Clause and the sections Construct 23 24 25 26 27 28 29 30 In the following example of the sections construct (Section 2.5.2 on page 48) the firstprivate clause is used to initialize the private copy of section_count of each thread. The problem is that the section constructs modify section_count, which breaks the independence of the section constructs. When different threads execute each section, both sections will print the value 1. When the same thread executes the two sections, one section will print the value 1 and the other will print the value 2. Since the order of execution of the two sections in this case is unspecified, it is unspecified which section prints which value. 31 Example A.13.1c 32 #include <omp.h> 21 A.13 190 OpenMP API • Version 3.1 July 2011 C/C++ 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 #include <stdio.h> #define NT 4 int main( ) { int section_count = 0; omp_set_dynamic(0); omp_set_num_threads(NT); #pragma omp parallel #pragma omp sections firstprivate( section_count ) { #pragma omp section { section_count++; /* may print the number one or two */ printf( "section_count %d\n", section_count ); } #pragma omp section { section_count++; /* may print the number one or two */ printf( "section_count %d\n", section_count ); } } return 1; } C/C++ Fortran 25 Example A.13.1f 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 program section use omp_lib integer :: section_count = 0 integer, parameter :: NT = 4 call omp_set_dynamic(.false.) call omp_set_num_threads(NT) !$omp parallel !$omp sections firstprivate ( section_count ) !$omp section section_count = section_count + 1 ! may print the number one or two print *, 'section_count', section_count !$omp section section_count = section_count + 1 ! may print the number one or two print *, 'section_count', section_count !$omp end sections !$omp end parallel end program section Fortran 45 Appendix A Examples 191 1 A.14 The single Construct 2 3 4 5 6 7 8 The following example demonstrates the single construct (Section 2.5.3 on page 50). In the example, only one thread prints each of the progress messages. All other threads will skip the single region and stop at the barrier at the end of the single construct until all threads in the team have reached the barrier. If other threads can proceed without waiting for the thread executing the single region, a nowait clause can be specified, as is done in the third single construct in this example. The user must not make any assumptions as to which thread will execute a single region. 9 Example A.14.1c 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 C/C++ #include <stdio.h> void work1() {} void work2() {} void single_example() { #pragma omp parallel { #pragma omp single printf("Beginning work1.\n"); work1(); #pragma omp single printf("Finishing work1.\n"); #pragma omp single nowait printf("Finished work1 and beginning work2.\n"); work2(); } } C/C++ 192 OpenMP API • Version 3.1 July 2011 Fortran Example A.14.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 SUBROUTINE WORK1() END SUBROUTINE WORK1 SUBROUTINE WORK2() END SUBROUTINE WORK2 PROGRAM SINGLE_EXAMPLE !$OMP PARALLEL !$OMP SINGLE print *, "Beginning work1." !$OMP END SINGLE CALL WORK1() !$OMP SINGLE print *, "Finishing work1." !$OMP END SINGLE !$OMP SINGLE print *, "Finished work1 and beginning work2." !$OMP END SINGLE NOWAIT CALL WORK2() !$OMP END PARALLEL END PROGRAM SINGLE_EXAMPLE Fortran 30 31 32 33 34 35 36 A.15 Tasking Constructs The following example shows how to traverse a tree-like structure using explicit tasks (see Section 2.7 on page 61). Note that the traverse function should be called from within a parallel region for the different specified tasks to be executed in parallel. Also note that the tasks will be executed in no specified order because there are no synchronization directives. Thus, assuming that the traversal will be done in post order, as in the sequential code, is wrong. Appendix A Examples 193 Example A.15.1c 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 C/C++ struct node { struct node *left; struct node *right; }; extern void process(struct node *); void traverse( struct node *p ) { if (p->left) #pragma omp task // p is firstprivate by default traverse(p->left); if (p->right) #pragma omp task // p is firstprivate by default traverse(p->right); process(p); } C/C++ Fortran Example A.15.1f 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 RECURSIVE SUBROUTINE traverse ( P ) TYPE Node TYPE(Node), POINTER :: left, right END TYPE Node TYPE(Node) :: P IF (associated(P%left)) THEN !$OMP TASK ! P is firstprivate by default call traverse(P%left) !$OMP END TASK ENDIF IF (associated(P%right)) THEN !$OMP TASK ! P is firstprivate by default call traverse(P%right) !$OMP END TASK ENDIF CALL process ( P ) END SUBROUTINE Fortran 194 OpenMP API • Version 3.1 July 2011 1 2 3 In the next example, we force a postorder traversal of the tree by adding a taskwait directive (see Section 2.8.4 on page 72). Now, we can safely assume that the left and right sons have been executed before we process the current node. 4 Example A.15.2c 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 C/C++ struct node { struct node *left; struct node *right; }; extern void process(struct node *); void postorder_traverse( struct node *p ) { if (p->left) #pragma omp task // p is firstprivate by default postorder_traverse(p->left); if (p->right) #pragma omp task // p is firstprivate by default postorder_traverse(p->right); #pragma omp taskwait process(p); } C/C++ Fortran 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 Example A.15.2f RECURSIVE SUBROUTINE traverse ( P ) TYPE Node TYPE(Node), POINTER :: left, right END TYPE Node TYPE(Node) :: P IF (associated(P%left)) THEN !$OMP TASK ! P is firstprivate by default call traverse(P%left) !$OMP END TASK ENDIF IF (associated(P%right)) THEN !$OMP TASK ! P is firstprivate by default call traverse(P%right) !$OMP END TASK ENDIF !$OMP TASKWAIT CALL process ( P ) END SUBROUTINE Fortran Appendix A Examples 195 1 2 3 4 5 The following example demonstrates how to use the task construct to process elements of a linked list in parallel. The thread executing the single region generates all of the explicit tasks, which are then executed by the threads in the current team. The pointer p is firstprivate by default on the task construct so it is not necessary to specify it in a firstprivate clause (see page 86). 6 Example A.15.3c 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 C/C++ typedef struct node node; struct node { int data; node * next; }; void process(node * p) { /* do work here */ } void increment_list_items(node * head) { #pragma omp parallel { #pragma omp single { node * p = head; while (p) { #pragma omp task // p is firstprivate by default process(p); p = p->next; } } } } C/C++ 196 OpenMP API • Version 3.1 July 2011 Fortran 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 Example A.15.3f MODULE LIST TYPE NODE INTEGER :: PAYLOAD TYPE (NODE), POINTER :: NEXT END TYPE NODE CONTAINS SUBROUTINE PROCESS(p) TYPE (NODE), POINTER :: P ! do work here END SUBROUTINE SUBROUTINE INCREMENT_LIST_ITEMS (HEAD) TYPE (NODE), POINTER :: HEAD TYPE (NODE), POINTER :: P !$OMP PARALLEL PRIVATE(P) !$OMP SINGLE P => HEAD DO !$OMP TASK ! P is firstprivate by default CALL PROCESS(P) !$OMP END TASK P => P%NEXT IF ( .NOT. ASSOCIATED (P) ) EXIT END DO !$OMP END SINGLE !$OMP END PARALLEL END SUBROUTINE END MODULE Fortran Appendix A Examples 197 1 2 3 4 The fib() function should be called from within a parallel region for the different specified tasks to be executed in parallel. Also, only one thread of the parallel region should call fib() unless multiple concurrent Fibonacci computations are desired. 5 Example A.15.4c 6 7 8 9 10 11 12 13 14 15 16 17 18 C/C++ int fib(int n) { int i, j; if (n<2) return n; else { #pragma omp task shared(i) i=fib(n-1); #pragma omp task shared(j) j=fib(n-2); #pragma omp taskwait return i+j; } } C/C++ Fortran Example A.15.4f 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 !$OMP !$OMP !$OMP !$OMP !$OMP RECURSIVE INTEGER FUNCTION fib(n) RESULT(res) INTEGER n, i, j IF ( n .LT. 2) THEN res = n ELSE TASK SHARED(i) i = fib( n-1 ) END TASK TASK SHARED(j) j = fib( n-2 ) END TASK TASKWAIT res = i+j END IF END FUNCTION Fortran Note: There are more efficient algorithms for computing Fibonacci numbers. This classic recursion algorithm is for illustrative purposes. 35 36 198 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 The following example demonstrates a way to generate a large number of tasks with one thread and execute them with the threads in the team (see Section 2.7.3 on page 65). While generating these tasks, the implementation may reach its limit on unassigned tasks. If it does, the implementation is allowed to cause the thread executing the task generating loop to suspend its task at the task scheduling point in the task directive, and start executing unassigned tasks. Once the number of unassigned tasks is sufficiently low, the thread may resume execution of the task generating loop. 8 Example A.15.5c 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 C/C++ #define LARGE_NUMBER 10000000 double item[LARGE_NUMBER]; extern void process(double); int main() { #pragma omp parallel { #pragma omp single { int i; for (i=0; i<LARGE_NUMBER; i++) #pragma omp task // i is firstprivate, item is shared process(item[i]); } } } C/C++ Fortran 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 Example A.15.5f real*8 item(10000000) integer i !$omp parallel !$omp single ! loop iteration variable i is private do i=1,10000000 !$omp task ! i is firstprivate, item is shared call process(item(i)) !$omp end task end do !$omp end single !$omp end parallel end Fortran Appendix A Examples 199 1 2 3 4 5 6 7 The following example is the same as the previous one, except that the tasks are generated in an untied task (see Section 2.7 on page 61). While generating the tasks, the implementation may reach its limit on unassigned tasks. If it does, the implementation is allowed to cause the thread executing the task generating loop to suspend its task at the task scheduling point in the task directive, and start executing unassigned tasks. If that thread begins execution of a task that takes a long time to complete, the other threads may complete all the other tasks before it is finished. 8 9 10 11 In this case, since the loop is in an untied task, any other thread is eligible to resume the task generating loop. In the previous examples, the other threads would be forced to idle until the generating thread finishes its long task, since the task generating loop was in a tied task. 12 Example A.15.6c 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 #define LARGE_NUMBER 10000000 double item[LARGE_NUMBER]; extern void process(double); int main() { #pragma omp parallel { #pragma omp single { int i; #pragma omp task untied // i is firstprivate, item is shared { for (i=0; i<LARGE_NUMBER; i++) #pragma omp task process(item[i]); } } } return 0; } C/C++ C/C++ 200 OpenMP API • Version 3.1 July 2011 Fortran 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Example A.15.6f real*8 item(10000000) !$omp parallel !$omp single !$omp task untied ! loop iteration variable i is private do i=1,10000000 !$omp task ! i is firstprivate, item is shared call process(item(i)) !$omp end task end do !$omp end task !$omp end single !$omp end parallel end Fortran 16 17 18 19 20 21 The following two examples demonstrate how the scheduling rules illustrated in Section 2.7.3 on page 65 affect the usage of threadprivate variables in tasks. A threadprivate variable can be modified by another task that is executed by the same thread. Thus, the value of a threadprivate variable cannot be assumed to be unchanged across a task scheduling point. In untied tasks, task scheduling points may be added in any place by the implementation. 22 23 24 A task switch may occur at a task scheduling point. A single thread may execute both of the task regions that modify tp. The parts of these task regions in which tp is modified may be executed in any order so the resulting value of var can be either 1 or 2. Appendix A Examples 201 Example A.15.7c 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 C/C++ int tp; #pragma omp threadprivate(tp) int var; void work() { #pragma omp task { /* do work here */ #pragma omp task { tp = 1; /* do work here */ #pragma omp task { /* no modification of tp */ } var = tp; //value of tp can be 1 or 2 } tp = 2; } } C/C++ Fortran Example A.15.7f 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 !$omp !$omp !$omp !$omp !$omp !$omp !$omp module example integer tp threadprivate(tp) integer var contains subroutine work use globals task ! do work here task tp = 1 ! do work here task ! no modification of tp end task var = tp ! value of var can be 1 or 2 end task tp = 2 end task end subroutine end module Fortran 202 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 In this example, scheduling constraints (see Section 2.7.3 on page 65) prohibit a thread in the team from executing a new task that modifies tp while another such task region tied to the same thread is suspended. Therefore, the value written will persist across the task scheduling point. Example A.15.8c C/C++ int tp; #pragma omp threadprivate(tp) int var; void work() { #pragma omp parallel { /* do work here */ #pragma omp task { tp++; /* do work here */ #pragma omp task { /* do work here but don't modify tp */ } var = tp; //Value does not change after write above } } } C/C++ Fortran 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 Example A.15.8f !$omp !$omp !$omp !$omp !$omp !$omp !$omp module example integer tp threadprivate(tp) integer var contains subroutine work parallel ! do work here task tp = tp + 1 ! do work here task ! do work here but don't modify tp end task var = tp ! value does not change after write above end task end parallel end subroutine Appendix A Examples 203 1 end module Fortran 2 3 4 5 The following two examples demonstrate how the scheduling rules illustrated in Section 2.7.3 on page 65 affect the usage of locks and critical sections in tasks. If a lock is held across a task scheduling point, no attempt should be made to acquire the same lock in any code that may be interleaved. Otherwise, a deadlock is possible. 6 7 8 In the example below, suppose the thread executing task 1 defers task 2. When it encounters the task scheduling point at task 3, it could suspend task 1 and begin task 2 which will result in a deadlock when it tries to enter critical region 1. 9 Example A.15.9c 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 C/C++ void work() { #pragma omp task { //Task 1 #pragma omp task { //Task 2 #pragma omp critical //Critical region 1 {/*do work here */ } } #pragma omp critical //Critical Region 2 { //Capture data for the following task #pragma omp task { /* do work here */ } //Task 3 } } } C/C++ 204 OpenMP API • Version 3.1 July 2011 Fortran 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Example A.15.9f !$omp !$omp !$omp !$omp !$omp !$omp !$omp !$omp !$omp !$omp module example contains subroutine work task ! Task 1 task ! Task 2 critical ! Critical region 1 ! do work here end critical end task critical ! Critical region 2 ! Capture data for the following task task !Task 3 ! do work here end task end critical end task end subroutine end module Fortran Appendix A Examples 205 1 2 3 4 In the following example, lock is held across a task scheduling point. However, according to the scheduling restrictions outlined in Section 2.7.3 on page 65, the executing thread can't begin executing one of the non-descendant tasks that also acquires lock before the task region is complete. Therefore, no deadlock is possible. 5 Example A.15.10c 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 C/C++ #include <omp.h> void work() { omp_lock_t lock; omp_init_lock(&lock); #pragma omp parallel { int i; #pragma omp for for (i = 0; i < 100; i++) { #pragma omp task { // lock is shared by default in the task omp_set_lock(&lock); // Capture data for the following task #pragma omp task // Task Scheduling Point 1 { /* do work here */ } omp_unset_lock(&lock); } } } omp_destroy_lock(&lock); } C/C++ 206 OpenMP API • Version 3.1 July 2011 Fortran 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 Example A.15.10f module example include 'omp_lib.h' integer (kind=omp_lock_kind) lock integer i contains subroutine work call omp_init_lock(lock) !$omp parallel !$omp do do i=1,100 !$omp task ! Outer task call omp_set_lock(lock) ! lock is shared by ! default in the task ! Capture data for the following task !$omp task ! Task Scheduling Point 1 ! do work here !$omp end task call omp_unset_lock(lock) !$omp end task end do !$omp end parallel call omp_destroy_lock(lock) end subroutine end module Fortran 27 28 29 30 31 32 33 34 35 36 The following examples illustrate the use of the mergeable clause in the task construct. In this first example, the task construct has been annotated with the mergeable clause (see Section 2.7.1 on page 61). The addition of this clause allows the implementation to reuse the data environment (including the ICVs) of the parent task for the task inside foo if the task is included or undeferred (see Section 1.2.3 on page 8). Thus, the result of the execution may differ depending on whether the task is merged or not. Therefore the mergeable clause needs to be used with caution. In this example, the use of the mergeable clause is safe. As x is a shared variable the outcome does not depend on whether or not the task is merged (that is, the task will always increment the same variable and will always compute the same value for x). 37 Example A.15.11c 38 39 40 41 #include <stdio.h> void foo ( ) { int x = 2; C/C++ Appendix A Examples 207 1 2 3 4 5 6 7 #pragma omp task shared(x) mergeable { x++; } #pragma omp taskwait printf("%d\n",x); // prints 3 } C/C++ Fortran Example A.15.11f 8 9 10 11 12 13 14 15 16 17 subroutine foo() integer :: x x = 2 !$omp task shared(x) mergeable x = x + 1 !$omp end task !$omp taskwait print *, x ! prints 3 end subroutine Fortran 18 19 20 21 22 This second example shows an incorrect use of the mergeable clause. In this example, the created task will access different instances of the variable x if the task is not merged, as x is firstprivate, but it will access the same variable x if the task is merged. As a result, the behavior of the program is unspecified and it can print two different values for x depending on the decisions taken by the implementation. 23 Example A.15.12c 24 25 26 27 28 29 30 31 32 33 34 #include <stdio.h> void foo ( ) { int x = 2; #pragma omp task mergeable { x++; } #pragma omp taskwait printf("%d\n",x); // prints 2 or 3 } C/C++ C/C++ 35 36 208 OpenMP API • Version 3.1 July 2011 Fortran 1 2 3 4 5 6 7 8 9 10 Example A.15.12f subroutine foo() integer :: x x = 2 !$omp task mergeable x = x + 1 !$omp end task !$omp taskwait print *, x ! prints 2 or 3 end subroutine Fortran 11 12 13 14 15 The following example shows the use of the final clause (see Section 2.7.1 on page 61) and the omp_in_final API call (see Section 3.2.20 on page 140) in a recursive binary search program. To reduce overhead, once a certain depth of recursion is reached the program uses the final clause to create only included tasks, which allow additional optimizations. 16 17 18 19 20 21 22 23 24 The use of the omp_in_final API call allows programmers to optimize their code by specifying which parts of the program are not necessary when a task can create only included tasks (that is, the code is inside a final task). In this example, the use of a different state variable is not necessary so once the program reaches the part of the computation that is finalized and copying from the parent state to the new state is eliminated. The allocation of new_state in the stack could also be avoided but it would make this example less clear. The final clause is most effective when used in conjunction with the mergeable clause since all tasks created in a final task region are included tasks that can be merged if the mergeable clause is present. 25 Example A.15.13c 26 27 28 29 30 31 32 33 34 35 36 37 38 39 #include <string.h> #include <omp.h> #define LIMIT 3 /* arbitrary limit on recursion depth */ void check_solution(char *); void bin_search (int pos, int n, char *state) { if ( pos == n ) { check_solution(state); return; } #pragma omp task final( pos > LIMIT ) mergeable { char new_state[n]; if (!omp_in_final() ) { C/C++ Appendix A Examples 209 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 memcpy(new_state, state, pos ); state = new_state; } state[pos] = 0; bin_search(pos+1, n, state ); } #pragma omp task final( pos > LIMIT ) mergeable { char new_state[n]; if (! omp_in_final() ) { memcpy(new_state, state, pos ); state = new_state; } state[pos] = 1; bin_search(pos+1, n, state ); } #pragma omp taskwait } C/C++ Fortran 19 Example A.15.13f 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 recursive subroutine bin_search(pos, n, state) use omp_lib integer :: pos, n character, pointer :: state(:) character, target, dimension(n) :: new_state1, new_state2 integer, parameter :: LIMIT = 3 if (pos .eq. n) then call check_solution(state) return endif !$omp task final(pos > LIMIT) mergeable if (.not. omp_in_final()) then new_state1(1:pos) = state(1:pos) state => new_state1 endif state(pos+1) = 'z' call bin_search(pos+1, n, state) !$omp end task !$omp task final(pos > LIMIT) mergeable if (.not. omp_in_final()) then new_state2(1:pos) = state(1:pos) state => new_state2 endif state(pos+1) = 'y' call bin_search(pos+1, n, state) !$omp end task !$omp taskwait 210 OpenMP API • Version 3.1 July 2011 1 end subroutine Fortran 2 3 4 5 6 7 8 The following example illustrates the difference between the if and the final clauses. The if clause has a local effect. In the first nest of tasks, the one that has the if clause will be undeferred but the task nested inside that task will not be affected by the if clause and will be created as usual. Alternatively, the final clause affects all task constructs in the final task region but not the final task itself. In the second nest of tasks, the nested tasks will be created as included tasks. Note also that the conditions for the if and final clauses are usually the opposite. 9 Example A.15.14c 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 C/C++ void foo ( ) { int i; #pragma omp task if(0) // This task is undeferred { #pragma omp task // This task is a regular task for (i = 0; i < 3; i++) { #pragma omp task // This task is a regular task bar(); } } #pragma omp task final(1) // This task is a regular task { #pragma omp task // This task is included for (i = 0; i < 3; i++) { #pragma omp task // This task is also included bar(); } } } C/C++ Fortran 30 Example A.15.14f 31 32 33 34 35 36 37 38 subroutine foo() integer i !$omp task if(.FALSE.) ! This task is undeferred !$omp task ! This task is a regular task do i = 1, 3 !$omp task ! This task is a regular task call bar() !$omp end task Appendix A Examples 211 1 2 3 4 5 6 7 8 9 10 11 12 13 enddo !$omp end task !$omp end task !$omp task final(.TRUE.) ! This task is a regular task !$omp task ! This task is included do i = 1, 3 !$omp task ! This task is also included call bar() !$omp end task enddo !$omp end task !$omp end task end subroutine Fortran 14 A.16 The taskyield Directive 15 16 17 18 19 The following example illustrates the use of the taskyield directive (see Section 2.7.2 on page 64). The tasks in the example compute something useful and then do some computation that must be done in a critical region. By using taskyield when a task cannot get access to the critical region the implementation can suspend the current task and schedule some other task that can do something useful. 20 Example A.16.1c 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 #include <omp.h> C/C++ void something_useful ( void ); void something_critical ( void ); void foo ( omp_lock_t * lock, int n ) { int i; for ( i = 0; i < n; i++ ) #pragma omp task { something_useful(); while ( !omp_test_lock(lock) ) { #pragma omp taskyield } something_critical(); omp_unset_lock(lock); } } C/C++ 212 OpenMP API • Version 3.1 July 2011 Fortran Example A.16.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 subroutine foo ( lock, n ) use omp_lib integer (kind=omp_lock_kind) :: lock integer n integer i do i = 1, n !$omp task call something_useful() do while ( .not. omp_test_lock(lock) ) !$omp taskyield end do call something_critical() call omp_unset_lock(lock) !$omp end task end do end subroutine Fortran 20 21 22 Fortran 23 24 A.17 The workshare Construct 25 26 The following are examples of the workshare construct (see Section 2.5.4 on page 52). 27 28 29 In the following example, workshare spreads work across the threads executing the parallel region, and there is a barrier after the last statement. Implementations must enforce Fortran execution rules inside of the workshare block. Appendix A Examples 213 Fortran (cont.) Example A.17.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 SUBROUTINE WSHARE1(AA, BB, CC, DD, EE, FF, N) INTEGER N REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N), EE(N,N), FF(N,N) !$OMP !$OMP !$OMP !$OMP PARALLEL WORKSHARE AA = BB CC = DD EE = FF END WORKSHARE END PARALLEL END SUBROUTINE WSHARE1 15 16 17 In the following example, the barrier at the end of the first workshare region is eliminated with a nowait clause. Threads doing CC = DD immediately begin work on EE = FF when they are done with CC = DD. 18 Example A.17.2f 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 SUBROUTINE WSHARE2(AA, BB, CC, DD, EE, FF, N) INTEGER N REAL AA(N,N), BB(N,N), CC(N,N) REAL DD(N,N), EE(N,N), FF(N,N) !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP PARALLEL WORKSHARE AA = BB CC = DD END WORKSHARE NOWAIT WORKSHARE EE = FF END WORKSHARE END PARALLEL END SUBROUTINE WSHARE2 The following example shows the use of an atomic directive inside a workshare construct. The computation of SUM(AA) is workshared, but the update to R is atomic. 34 35 214 OpenMP API • Version 3.1 July 2011 Fortran (cont.) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Example A.17.3f SUBROUTINE WSHARE3(AA, BB, CC, DD, N) INTEGER N REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N) REAL R !$OMP !$OMP !$OMP !$OMP !$OMP R=0 PARALLEL WORKSHARE AA = BB ATOMIC UPDATE R = R + SUM(AA) CC = DD END WORKSHARE END PARALLEL END SUBROUTINE WSHARE3 18 19 20 21 Fortran WHERE and FORALL statements are compound statements, made up of a control part and a statement part. When workshare is applied to one of these compound statements, both the control and the statement parts are workshared. The following example shows the use of a WHERE statement in a workshare construct. 22 23 24 25 26 27 Each task gets worked on in order by the threads: AA = BB then CC = DD then EE .ne. 0 then FF = 1 / EE then GG = HH 28 Example A.17.4f 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 SUBROUTINE WSHARE4(AA, BB, CC, DD, EE, FF, GG, HH, N) INTEGER N REAL AA(N,N), BB(N,N), CC(N,N) REAL DD(N,N), EE(N,N), FF(N,N) REAL GG(N,N), HH(N,N) !$OMP !$OMP !$OMP !$OMP PARALLEL WORKSHARE AA = BB CC = DD WHERE (EE .ne. 0) FF = 1 / EE GG = HH END WORKSHARE END PARALLEL END SUBROUTINE WSHARE4 Appendix A Examples 215 Fortran (cont.) 1 2 In the following example, an assignment to a shared scalar variable is performed by one thread in a workshare while all other threads in the team wait. 3 Example A.17.5f 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 SUBROUTINE WSHARE5(AA, BB, CC, DD, N) INTEGER N REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N) INTEGER SHR !$OMP !$OMP !$OMP !$OMP PARALLEL SHARED(SHR) WORKSHARE AA = BB SHR = 1 CC = DD * SHR END WORKSHARE END PARALLEL END SUBROUTINE WSHARE5 20 21 22 23 The following example contains an assignment to a private scalar variable, which is performed by one thread in a workshare while all other threads wait. It is nonconforming because the private scalar variable is undefined after the assignment statement. 24 Example A.17.6f 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 SUBROUTINE WSHARE6_WRONG(AA, BB, CC, DD, N) INTEGER N REAL AA(N,N), BB(N,N), CC(N,N), DD(N,N) INTEGER PRI !$OMP !$OMP !$OMP !$OMP PARALLEL PRIVATE(PRI) WORKSHARE AA = BB PRI = 1 CC = DD * PRI END WORKSHARE END PARALLEL END SUBROUTINE WSHARE6_WRONG 216 OpenMP API • Version 3.1 July 2011 1 2 3 4 Fortran execution rules must be enforced inside a workshare construct. In the following example, the same result is produced in the following program fragment regardless of whether the code is executed sequentially or inside an OpenMP program with multiple threads: 5 Example A.17.7f 6 7 8 9 10 11 12 13 14 15 16 17 SUBROUTINE WSHARE7(AA, BB, CC, N) INTEGER N REAL AA(N), BB(N), CC(N) !$OMP !$OMP !$OMP !$OMP PARALLEL WORKSHARE AA(1:50) = BB(11:60) CC(11:20) = AA(1:10) END WORKSHARE END PARALLEL END SUBROUTINE WSHARE7 Fortran 18 A.18 The master Construct 19 20 21 The following example demonstrates the master construct (Section 2.8.1 on page 67). In the example, the master keeps track of how many iterations have been executed and prints out a progress report. The other threads skip the master region without waiting. 22 Example A.18.1c 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 #include <stdio.h> C/C++ extern float average(float,float,float); void master_example( float* x, float* xold, int n, float tol ) { int c, i, toobig; float error, y; c = 0; #pragma omp parallel { do{ #pragma omp for private(i) for( i = 1; i < n-1; ++i ){ xold[i] = x[i]; Appendix A Examples 217 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 } #pragma omp single { toobig = 0; } #pragma omp for private(i,y,error) reduction(+:toobig) for( i = 1; i < n-1; ++i ){ y = x[i]; x[i] = average( xold[i-1], x[i], xold[i+1] ); error = y - x[i]; if( error > tol || error < -tol ) ++toobig; } #pragma omp master { ++c; printf( "iteration %d, toobig=%d\n", c, toobig ); } }while( toobig > 0 ); } } C/C++ 218 OpenMP API • Version 3.1 July 2011 Fortran Example A.18.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP SUBROUTINE MASTER_EXAMPLE( X, XOLD, N, TOL ) REAL X(*), XOLD(*), TOL INTEGER N INTEGER C, I, TOOBIG REAL ERROR, Y, AVERAGE EXTERNAL AVERAGE C = 0 TOOBIG = 1 PARALLEL DO WHILE( TOOBIG > 0 ) DO PRIVATE(I) DO I = 2, N-1 XOLD(I) = X(I) ENDDO SINGLE TOOBIG = 0 END SINGLE DO PRIVATE(I,Y,ERROR), REDUCTION(+:TOOBIG) DO I = 2, N-1 Y = X(I) X(I) = AVERAGE( XOLD(I-1), X(I), XOLD(I+1) ) ERROR = Y-X(I) IF( ERROR > TOL .OR. ERROR < -TOL ) TOOBIG = TOOBIG+1 ENDDO MASTER C = C + 1 PRINT *, 'Iteration ', C, 'TOOBIG=', TOOBIG END MASTER ENDDO END PARALLEL END SUBROUTINE MASTER_EXAMPLE Fortran 33 34 35 36 37 38 39 A.19 The critical Construct The following example includes several critical constructs (Section 2.8.2 on page 68). The example illustrates a queuing model in which a task is dequeued and worked on. To guard against multiple threads dequeuing the same task, the dequeuing operation must be in a critical region. Because the two queues in this example are independent, they are protected by critical constructs with different names, xaxis and yaxis. Appendix A Examples 219 Example A.19.1c 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 C/C++ int dequeue(float *a); void work(int i, float *a); void critical_example(float *x, float *y) { int ix_next, iy_next; #pragma omp parallel shared(x, y) private(ix_next, iy_next) { #pragma omp critical (xaxis) ix_next = dequeue(x); work(ix_next, x); #pragma omp critical (yaxis) iy_next = dequeue(y); work(iy_next, y); } } C/C++ Fortran Example A.19.1f 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 SUBROUTINE CRITICAL_EXAMPLE(X, Y) REAL X(*), Y(*) INTEGER IX_NEXT, IY_NEXT !$OMP PARALLEL SHARED(X, Y) PRIVATE(IX_NEXT, IY_NEXT) !$OMP CRITICAL(XAXIS) CALL DEQUEUE(IX_NEXT, X) !$OMP END CRITICAL(XAXIS) CALL WORK(IX_NEXT, X) !$OMP CRITICAL(YAXIS) CALL DEQUEUE(IY_NEXT,Y) !$OMP END CRITICAL(YAXIS) CALL WORK(IY_NEXT, Y) !$OMP END PARALLEL END SUBROUTINE CRITICAL_EXAMPLE Fortran 220 OpenMP API • Version 3.1 July 2011 1 2 A.20 worksharing Constructs Inside a critical Construct 3 4 5 6 7 8 9 10 The following example demonstrates using a worksharing construct inside a critical construct (see Section 2.8.2 on page 68). This example is conforming because the worksharing single region is not closely nested inside the critical region (see Section 2.10 on page 111). A single thread executes the one and only section in the sections region, and executes the critical region. The same thread encounters the nested parallel region, creates a new team of threads, and becomes the master of the new team. One of the threads in the new team enters the single region and increments i by 1. At the end of this example i is equal to 2. 11 Example A.20.1c 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 void critical_work() { int i = 1; #pragma omp parallel sections { #pragma omp section { #pragma omp critical (name) { #pragma omp parallel { #pragma omp single { i++; } } } } } } C/C++ C/C++ Appendix A Examples 221 Fortran Example A.20.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 SUBROUTINE CRITICAL_WORK() INTEGER I I = 1 !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP PARALLEL SECTIONS SECTION CRITICAL (NAME) PARALLEL SINGLE I = I + 1 END SINGLE END PARALLEL END CRITICAL (NAME) END PARALLEL SECTIONS END SUBROUTINE CRITICAL_WORK Fortran 18 A.21 Binding of barrier Regions 19 20 The binding rules call for a barrier region to bind to the closest enclosing parallel region (see Section 2.8.3 on page 70). 21 22 23 24 In the following example, the call from the main program to sub2 is conforming because the barrier region (in sub3) binds to the parallel region in sub2. The call from the main program to sub1 is conforming because the barrier region binds to the parallel region in subroutine sub2. 25 26 27 28 The call from the main program to sub3 is conforming because the barrier region binds to the implicit inactive parallel region enclosing the sequential part. Also note that the barrier region in sub3 when called from sub2 only synchronizes the team of threads in the enclosing parallel region and not all the threads created in sub1. 222 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 Example A.21.1c C/C++ void work(int n) {} void sub3(int n) { work(n); #pragma omp barrier work(n); } void sub2(int k) { #pragma omp parallel shared(k) sub3(k); } void sub1(int n) { int i; #pragma omp parallel private(i) shared(n) { #pragma omp for for (i=0; i<n; i++) sub2(i); } } int main() { sub1(2); sub2(2); sub3(2); return 0; } C/C++ Appendix A Examples 223 Fortran Example A.21.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 SUBROUTINE WORK(N) INTEGER N END SUBROUTINE WORK SUBROUTINE SUB3(N) INTEGER N CALL WORK(N) !$OMP BARRIER CALL WORK(N) END SUBROUTINE SUB3 SUBROUTINE SUB2(K) INTEGER K !$OMP PARALLEL SHARED(K) CALL SUB3(K) !$OMP END PARALLEL END SUBROUTINE SUB2 SUBROUTINE SUB1(N) INTEGER N INTEGER I !$OMP PARALLEL PRIVATE(I) SHARED(N) !$OMP DO DO I = 1, N CALL SUB2(I) END DO !$OMP END PARALLEL END SUBROUTINE SUB1 PROGRAM EXAMPLE CALL SUB1(2) CALL SUB2(2) CALL SUB3(2) END PROGRAM EXAMPLE Fortran 37 A.22 The atomic Construct The following example avoids race conditions (simultaneous updates of an element of x by multiple threads) by using the atomic construct (Section 2.8.5 on page 73). 38 39 224 OpenMP API • Version 3.1 July 2011 1 2 3 4 The advantage of using the atomic construct in this example is that it allows updates of two different elements of x to occur in parallel. If a critical construct (see Section 2.8.2 on page 68) were used instead, then all updates to elements of x would be executed serially (though not in any guaranteed order). 5 6 Note that the atomic directive applies only to the statement immediately following it. As a result, elements of y are not updated atomically in this example. 7 Example A.22.1c 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 C/C++ float work1(int i) { return 1.0 * i; } float work2(int i) { return 2.0 * i; } void atomic_example(float *x, float *y, int *index, int n) { int i; #pragma omp parallel for shared(x, y, index, n) for (i=0; i<n; i++) { #pragma omp atomic update x[index[i]] += work1(i); y[i] += work2(i); } } int main() { float x[1000]; float y[10000]; int index[10000]; int i; for (i = 0; i < 10000; i++) { index[i] = i % 1000; y[i]=0.0; } for (i = 0; i < 1000; i++) x[i] = 0.0; atomic_example(x, y, index, 10000); return 0; } C/C++ Appendix A Examples 225 Fortran Example A.22.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 REAL FUNCTION WORK1(I) INTEGER I WORK1 = 1.0 * I RETURN END FUNCTION WORK1 REAL FUNCTION WORK2(I) INTEGER I WORK2 = 2.0 * I RETURN END FUNCTION WORK2 SUBROUTINE SUB(X, Y, INDEX, N) REAL X(*), Y(*) INTEGER INDEX(*), N INTEGER I !$OMP !$OMP PARALLEL DO SHARED(X, Y, INDEX, N) DO I=1,N ATOMIC UPDATE X(INDEX(I)) = X(INDEX(I)) + WORK1(I) Y(I) = Y(I) + WORK2(I) ENDDO END SUBROUTINE SUB PROGRAM ATOMIC_EXAMPLE REAL X(1000), Y(10000) INTEGER INDEX(10000) INTEGER I DO I=1,10000 INDEX(I) = MOD(I, 1000) + 1 Y(I) = 0.0 ENDDO DO I = 1,1000 X(I) = 0.0 ENDDO CALL SUB(X, Y, INDEX, 10000) END PROGRAM ATOMIC_EXAMPLE Fortran 226 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 The following example illustrates the read and write clauses for the atomic directive. These clauses ensure that the given variable is read or written, respectively, as a whole. Otherwise, some other thread might read or write part of the variable while the current thread was reading or writing another part of the variable. Note that most hardware provides atomic reads and writes for some set of properly aligned variables of specific sizes, but not necessarily for all the variable types supported by the OpenMP API. 8 Example A.22.2c 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 C/C++ int atomic_read(const int *p) { int value; /* Guarantee that the entire value of *p is read atomically. No part of * *p can change during the read operation. */ #pragma omp atomic read value = *p; return value; } void atomic_write(int *p, int value) { /* Guarantee that value is stored atomically into *p. No part of *p can change * until after the entire write operation is completed. */ #pragma omp atomic write *p = value; } C/C++ Fortran 28 Example A.22.2f 29 30 31 32 33 34 35 36 37 38 39 40 41 function atomic_read(p) integer :: atomic_read integer, intent(in) :: p ! Guarantee that the entire value of p is read atomically. No part of ! p can change during the read operation. !$omp atomic read atomic_read = p return end function atomic_read subroutine atomic_write(p, value) integer, intent(out) :: p Appendix A Examples 227 1 2 3 4 5 6 integer, intent(in) :: value ! Guarantee that value is stored atomically into p. No part of p can change ! until after the entire write operation is completed. !$omp atomic write p = value end subroutine atomic_write Fortran 7 8 9 10 11 12 The following example illustrates the capture clause for the atomic directive. In this case the value of a variable is captured, and then the variable is incremented. These operations occur atomically. This particular example could be implemented using the fetch-and-add instruction available on many kinds of hardware. The example also shows a way to implement a spin lock using the capture and read clauses. 13 Example A.22.3c 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 int fetch_and_add(int *p) { /* Atomically read the value of *p and then increment it. The previous value is * returned. This can be used to implement a simple lock as shown below. */ int old; #pragma omp atomic capture { old = *p; (*p)++; } return old; } C/C++ /* * Use fetch_and_add to implement a lock */ struct locktype { int ticketnumber; int turn; }; void do_locked_work(struct locktype *lock) { int atomic_read(const int *p); void work(); // Obtain the lock int myturn = fetch_and_add(&lock->ticketnumber); while (atomic_read(&lock->turn) != myturn) ; // Do some work. The flush is needed to ensure visibility of // variables not involved in atomic directives #pragma omp flush work(); 228 OpenMP API • Version 3.1 July 2011 1 2 3 4 #pragma omp flush // Release the lock fetch_and_add(&lock->turn); } C/C++ Fortran 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 Example A.22.3f function fetch_and_add(p) integer:: fetch_and_add integer, intent(inout) :: p ! Atomically read the value of p and then increment it. The previous value is ! returned. This can be used to implement a simple lock as shown below. !$omp atomic capture fetch_and_add = p p = p + 1 !$omp end atomic end function fetch_and_add ! Use fetch_and_add to implement a lock module m interface function fetch_and_add(p) integer :: fetch_and_add integer, intent(inout) :: p end function function atomic_read(p) integer :: atomic_read integer, intent(in) :: p end function end interface type locktype integer ticketnumber integer turn end type contains subroutine do_locked_work(lock) type(locktype), intent(inout) :: lock integer myturn integer junk ! obtain the lock myturn = fetch_and_add(lock%ticketnumber) do while (atomic_read(lock%turn) .ne. myturn) continue enddo ! Do some work. The flush is needed to ensure visibility of variables ! not involved in atomic directives Appendix A Examples 229 1 2 3 4 5 6 7 8 !$omp flush call work !$omp flush ! Release the lock junk = fetch_and_add(lock%turn) end subroutine end module Fortran 9 10 11 12 A.23 Restrictions on the atomic Construct 13 14 The following non-conforming examples illustrate the restrictions on the atomic construct given in Section 2.8.5 on page 73. 15 Example A.23.1c 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 void atomic_wrong () { union {int n; float x;} u; C/C++ #pragma omp parallel { #pragma omp atomic update u.n++; #pragma omp atomic update u.x += 1.0; /* Incorrect because the atomic constructs reference the same location through incompatible types */ } } C/C++ Fortran Example A.23.1f 32 33 34 SUBROUTINE ATOMIC_WRONG() INTEGER:: I 230 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 REAL:: R EQUIVALENCE(I,R) !$OMP !$OMP PARALLEL ATOMIC UPDATE I = I + 1 !$OMP ATOMIC UPDATE R = R + 1.0 ! incorrect because I and R reference the same location ! but have different types !$OMP END PARALLEL END SUBROUTINE ATOMIC_WRONG Fortran 13 Example A.23.2c 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 void atomic_wrong2 () { int x; int *i; float *r; C/C++ i = &x; r = (float *)&x; #pragma omp parallel { #pragma omp atomic update *i += 1; #pragma omp atomic update *r += 1.0; /* Incorrect because the atomic constructs reference the same location through incompatible types */ } } C/C++ Appendix A Examples 231 Fortran 1 2 The following example is non-conforming because I and R reference the same location but have different types. 3 Example A.23.2f 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 SUBROUTINE SUB() COMMON /BLK/ R REAL R !$OMP ATOMIC UPDATE R = R + 1.0 END SUBROUTINE SUB SUBROUTINE ATOMIC_WRONG2() COMMON /BLK/ I INTEGER I !$OMP PARALLEL !$OMP ATOMIC UPDATE I = I + 1 CALL SUB() !$OMP END PARALLEL END SUBROUTINE ATOMIC_WRONG2 232 OpenMP API • Version 3.1 July 2011 1 2 Although the following example might work on some implementations, this is also nonconforming: 3 Example A.23.3f 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 SUBROUTINE ATOMIC_WRONG3 INTEGER:: I REAL:: R EQUIVALENCE(I,R) !$OMP !$OMP PARALLEL ATOMIC UPDATE I = I + 1 ! incorrect because I and R reference the same location ! but have different types !$OMP END PARALLEL !$OMP !$OMP PARALLEL ATOMIC UPDATE R = R + 1.0 ! incorrect because I and R reference the same location ! but have different types !$OMP END PARALLEL END SUBROUTINE ATOMIC_WRONG3 Fortran 24 A.24 The flush Construct without a List 25 26 The following example (for Section 2.8.6 on page 78) distinguishes the shared variables affected by a flush construct with no list from the shared objects that are not affected: 27 Example A.24.1c 28 29 30 31 32 33 34 35 36 37 int x, *p = &x; C/C++ void f1(int *q) { *q = 1; #pragma omp flush /* x, p, and *q are flushed */ /* because they are shared and accessible */ /* q is not flushed because it is not shared. */ } Appendix A Examples 233 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 void f2(int *q) { #pragma omp barrier *q = 2; #pragma omp barrier /* /* /* /* a barrier implies a flush */ x, p, and *q are flushed */ because they are shared and accessible */ q is not flushed because it is not shared. */ } int g(int n) { int i = 1, j, sum = 0; *p = 1; #pragma omp parallel reduction(+: sum) num_threads(10) { f1(&j); /* i, n and sum were not flushed */ /* because they were not accessible in f1 */ /* j was flushed because it was accessible */ sum += j; f2(&j); /* i, n, and sum were not flushed */ /* because they were not accessible in f2 */ /* j was flushed because it was accessible */ sum += i + j + *p + n; } return sum; } int main() { int result = g(7); return result; } C/C++ 234 OpenMP API • Version 3.1 July 2011 Fortran 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 Example A.24.1f SUBROUTINE F1(Q) COMMON /DATA/ X, P INTEGER, TARGET :: X INTEGER, POINTER :: P INTEGER Q !$OMP Q = 1 FLUSH ! X, P and Q are flushed ! because they are shared and accessible END SUBROUTINE F1 SUBROUTINE F2(Q) COMMON /DATA/ X, P INTEGER, TARGET :: X INTEGER, POINTER :: P INTEGER Q !$OMP BARRIER Q = 2 !$OMP BARRIER ! a barrier implies a flush ! X, P and Q are flushed ! because they are shared and accessible END SUBROUTINE F2 INTEGER FUNCTION G(N) COMMON /DATA/ X, P INTEGER, TARGET :: X INTEGER, POINTER :: P INTEGER N INTEGER I, J, SUM !$OMP !$OMP I = 1 SUM = 0 P = 1 PARALLEL REDUCTION(+: SUM) NUM_THREADS(10) CALL F1(J) ! I, N and SUM were not flushed ! because they were not accessible in F1 ! J was flushed because it was accessible SUM = SUM + J CALL F2(J) ! I, N, and SUM were not flushed ! because they were not accessible in f2 ! J was flushed because it was accessible SUM = SUM + I + J + P + N END PARALLEL Appendix A Examples 235 1 2 3 4 5 6 7 8 9 10 11 12 13 14 G = SUM END FUNCTION G PROGRAM FLUSH_NOLIST COMMON /DATA/ X, P INTEGER, TARGET :: X INTEGER, POINTER :: P INTEGER RESULT, G P => X RESULT = G(7) PRINT *, RESULT END PROGRAM FLUSH_NOLIST Fortran 15 17 Placement of flush, barrier, taskwait and taskyield Directives 18 19 20 21 The following example is non-conforming, because the flush, barrier, taskwait, and taskyield directives are stand-alone directives and cannot be the immediate substatement of an if statement. See Section 2.8.3 on page 70, Section 2.8.6 on page 78, Section 2.8.4 on page 72, and Section 2.7.2 on page 64. 22 Example A.25.1c 16 A.25 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 C/C++ void standalone_wrong() { int a = 1; if (a != 0) #pragma omp flush(a) /* incorrect as flush cannot be immediate substatement of if statement */ if (a != 0) #pragma omp barrier /* incorrect as barrier cannot be immediate substatement of if statement */ if (a!=0) #pragma omp taskyield /* incorrect as taskyield cannot be immediate substatement of if statement */ 236 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 if (a != 0) #pragma omp taskwait /* incorrect as taskwait cannot be immediate substatement of if statement */ } C/C++ 7 8 9 10 The following example is non-conforming, because the flush, barrier, taskwait, and taskyield directives are stand-alone directives and cannot be the action statement of an if statement or a labeled branch target. Fortran 11 Example A.25.1f 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 SUBROUTINE STANDALONE_WRONG() INTEGER A A = 1 ! the FLUSH directive must not be the action statement ! in an IF statement IF (A .NE. 0) !$OMP FLUSH(A) ! the BARRIER directive must not be the action statement ! in an IF statement IF (A .NE. 0) !$OMP BARRIER ! the TASKWAIT directive must not be the action statement ! in an IF statement IF (A .NE. 0) !$OMP TASKWAIT ! the TASKYIELD directive must not be the action statement ! in an IF statement IF (A .NE. 0) !$OMP TASKYIELD GOTO 100 ! the FLUSH directive must not be a labeled branch target ! statement 100 !$OMP FLUSH(A) GOTO 200 ! the BARRIER directive must not be a labeled branch target ! statement 200 !$OMP BARRIER GOTO 300 ! the TASKWAIT directive must not be a labeled branch target ! statement 300 !$OMP TASKWAIT Appendix A Examples 237 1 2 3 4 5 6 7 GOTO 400 ! the TASKYIELD directive must not be a labeled branch target ! statement 400 !$OMP TASKYIELD END SUBROUTINE Fortran 8 9 10 The following version of the above example is conforming because the flush, barrier, taskwait, and taskyield directives are enclosed in a compound statement. 11 Example A.25.2c 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 void standalone_ok() { int a = 1; #pragma omp { if (a != #pragma omp } if (a != #pragma omp } if (a != #pragma omp } if (a != 0) #pragma omp } } C/C++ parallel 0) { flush(a) 0) { barrier 0) { taskwait { taskyield } C/C++ The following example is conforming because the flush, barrier, taskwait, and taskyield directives are enclosed in an if construct or follow the labeled branch target. 32 33 34 Fortran 35 Example A.25.2f 36 37 38 SUBROUTINE STANDALONE_OK() INTEGER A A = 1 238 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 IF (A .NE. 0) THEN !$OMP FLUSH(A) ENDIF IF (A .NE. 0) THEN !$OMP BARRIER ENDIF IF (A .NE. 0) THEN !$OMP TASKWAIT ENDIF IF (A .NE. 0) THEN !$OMP TASKYIELD ENDIF GOTO 100 100 CONTINUE !$OMP FLUSH(A) GOTO 200 200 CONTINUE !$OMP BARRIER GOTO 300 300 CONTINUE !$OMP TASKWAIT GOTO 400 400 CONTINUE !$OMP TASKYIELD END SUBROUTINE Fortran 26 28 The ordered Clause and the ordered Construct 29 30 31 Ordered constructs (Section 2.8.7 on page 82) are useful for sequentially ordering the output from work that is done in parallel. The following program prints out the indices in sequential order: 27 A.26 Appendix A Examples 239 Example A.26.1c 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 C/C++ #include <stdio.h> void work(int k) { #pragma omp ordered printf(" %d\n", k); } void ordered_example(int lb, int ub, int stride) { int i; #pragma omp parallel for ordered schedule(dynamic) for (i=lb; i<ub; i+=stride) work(i); } int main() { ordered_example(0, 100, 5); return 0; } C/C++ 240 OpenMP API • Version 3.1 July 2011 Fortran 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Example A.26.1f SUBROUTINE WORK(K) INTEGER k !$OMP ORDERED WRITE(*,*) K !$OMP END ORDERED END SUBROUTINE WORK SUBROUTINE SUB(LB, UB, STRIDE) INTEGER LB, UB, STRIDE INTEGER I !$OMP PARALLEL DO ORDERED SCHEDULE(DYNAMIC) DO I=LB,UB,STRIDE CALL WORK(I) END DO !$OMP END PARALLEL DO END SUBROUTINE SUB PROGRAM ORDERED_EXAMPLE CALL SUB(1,100,5) END PROGRAM ORDERED_EXAMPLE Fortran 26 27 28 29 It is possible to have multiple ordered constructs within a loop region with the ordered clause specified. The first example is non-conforming because all iterations execute two ordered regions. An iteration of a loop must not execute more than one ordered region: Appendix A Examples 241 C/C++ Example A.26.2c 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 void work(int i) {} void ordered_wrong(int n) { int i; #pragma omp for ordered for (i=0; i<n; i++) { /* incorrect because an iteration may not execute more than one ordered region */ #pragma omp ordered work(i); #pragma omp ordered work(i+1); } } C/C++ Fortran Example A.26.2f 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 SUBROUTINE WORK(I) INTEGER I END SUBROUTINE WORK SUBROUTINE ORDERED_WRONG(N) INTEGER N INTEGER I DO ORDERED DO I = 1, N ! incorrect because an iteration may not execute more than one ! ordered region !$OMP ORDERED CALL WORK(I) !$OMP END ORDERED !$OMP !$OMP ORDERED CALL WORK(I+1) !$OMP END ORDERED END DO END SUBROUTINE ORDERED_WRONG Fortran 242 OpenMP API • Version 3.1 July 2011 1 2 The following is a conforming example with more than one ordered construct. Each iteration will execute only one ordered region: 3 Example A.26.3c 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 C/C++ void work(int i) {} void ordered_good(int n) { int i; #pragma omp for ordered for (i=0; i<n; i++) { if (i <= 10) { #pragma omp ordered work(i); } if (i > 10) { #pragma omp ordered work(i+1); } } } C/C++ Fortran 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 Example A.26.3f SUBROUTINE ORDERED_GOOD(N) INTEGER N !$OMP !$OMP !$OMP DO ORDERED DO I = 1,N IF (I <= 10) THEN ORDERED CALL WORK(I) END ORDERED ENDIF IF (I > 10) THEN ORDERED CALL WORK(I+1) !$OMP END ORDERED ENDIF ENDDO END SUBROUTINE ORDERED_GOOD !$OMP Fortran Appendix A Examples 243 1 A.27 The threadprivate Directive 2 3 The following examples demonstrate how to use the threadprivate directive (Section 2.9.2 on page 88) to give each thread a separate counter. 4 Example A.27.1c 5 6 7 8 9 10 11 12 C/C++ int counter = 0; #pragma omp threadprivate(counter) int increment_counter() { counter++; return(counter); } C/C++ Fortran 13 Example A.27.1f 14 15 16 17 18 19 20 21 INTEGER FUNCTION INCREMENT_COUNTER() COMMON/INC_COMMON/COUNTER !$OMP THREADPRIVATE(/INC_COMMON/) COUNTER = COUNTER +1 INCREMENT_COUNTER = COUNTER RETURN END FUNCTION INCREMENT_COUNTER Fortran C/C++ 22 The following example uses threadprivate on a static variable: 23 Example A.27.2c 24 25 26 27 28 29 30 int increment_counter_2() { static int counter = 0; #pragma omp threadprivate(counter) counter++; return(counter); } 244 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 The following example demonstrates unspecified behavior for the initialization of a threadprivate variable. A threadprivate variable is initialized once at an unspecified point before its first reference. Because a is constructed using the value of x (which is modified by the statement x++), the value of a.val at the start of the parallel region could be either 1 or 2. This problem is avoided for b, which uses an auxiliary const variable and a copy-constructor. 8 Example A.27.3c 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 class T { public: int val; T (int); T (const T&); }; T :: T (int v){ val = v; } T :: T (const T& t) { val = t.val; } void g(T a, T b){ a.val += b.val; } int x = 1; T a(x); const T b_aux(x); /* Capture value of x = 1 */ T b(b_aux); #pragma omp threadprivate(a, b) void f(int n) { x++; #pragma omp parallel for /* In each thread: * a is constructed from x (with value 1 or 2?) * b is copy-constructed from b_aux */ for (int i=0; i<n; i++) { g(a, b); /* Value of a is unspecified. */ } } C/C++ Appendix A Examples 245 Fortran 1 2 3 The following examples show non-conforming uses and correct uses of the threadprivate directive. For more information, see Section 2.9.2 on page 88 and Section 2.9.4.1 on page 107. 4 5 The following example is non-conforming because the common block is not declared local to the subroutine that refers to it: 6 Example A.27.2f 7 8 9 10 11 12 13 14 15 MODULE INC_MODULE COMMON /T/ A END MODULE INC_MODULE SUBROUTINE INC_MODULE_WRONG() USE INC_MODULE !$OMP THREADPRIVATE(/T/) !non-conforming because /T/ not declared in INC_MODULE_WRONG END SUBROUTINE INC_MODULE_WRONG 16 17 18 The following example is also non-conforming because the common block is not declared local to the subroutine that refers to it: 19 Example A.27.3f 20 21 22 23 24 25 26 27 28 29 30 31 SUBROUTINE INC_WRONG() COMMON /T/ A !$OMP THREADPRIVATE(/T/) CONTAINS SUBROUTINE INC_WRONG_SUB() !$OMP PARALLEL COPYIN(/T/) !non-conforming because /T/ not declared in INC_WRONG_SUB !$OMP END PARALLEL END SUBROUTINE INC_WRONG_SUB END SUBROUTINE INC_WRONG 246 OpenMP API • Version 3.1 July 2011 Fortran (cont.) 1 The following example is a correct rewrite of the previous example: 2 Example A.27.4f 3 4 5 6 7 8 9 10 11 12 13 14 15 16 !$OMP !$OMP !$OMP !$OMP SUBROUTINE INC_GOOD() COMMON /T/ A THREADPRIVATE(/T/) CONTAINS SUBROUTINE INC_GOOD_SUB() COMMON /T/ A THREADPRIVATE(/T/) PARALLEL COPYIN(/T/) END PARALLEL END SUBROUTINE INC_GOOD_SUB END SUBROUTINE INC_GOOD 17 The following is an example of the use of threadprivate for local variables: 18 Example A.27.5f 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 PROGRAM INC_GOOD2 INTEGER, ALLOCATABLE, SAVE :: A(:) INTEGER, POINTER, SAVE :: PTR INTEGER, SAVE :: I INTEGER, TARGET :: TARG LOGICAL :: FIRSTIN = .TRUE. !$OMP THREADPRIVATE(A, I, PTR) ALLOCATE (A(3)) A = (/1,2,3/) PTR => TARG I = 5 !$OMP !$OMP PARALLEL COPYIN(I, PTR) CRITICAL IF (FIRSTIN) THEN TARG = 4 ! Update target of ptr I = I + 10 IF (ALLOCATED(A)) A = A + 10 FIRSTIN = .FALSE. END IF IF (ALLOCATED(A)) THEN PRINT *, 'a = ', A ELSE Appendix A Examples 247 Fortran (cont.) 1 2 3 4 5 6 7 8 9 10 11 PRINT *, 'A is not allocated' END IF PRINT *, 'ptr = ', PTR PRINT *, 'i = ', I PRINT * !$OMP !$OMP END CRITICAL END PARALLEL END PROGRAM INC_GOOD2 The above program, if executed by two threads, will print one of the following two sets of output: 12 13 14 15 16 17 18 19 20 21 a = 11 12 13 ptr = 4 i = 15 A is not allocated ptr = 4 i = 5 or 22 23 24 25 26 27 28 29 30 A is not allocated ptr = 4 i = 15 a = 1 2 3 ptr = 4 i = 5 31 32 The following is an example of the use of threadprivate for module variables: 33 Example A.27.6f 34 35 36 37 38 39 40 41 42 43 MODULE INC_MODULE_GOOD3 REAL, POINTER :: WORK(:) SAVE WORK !$OMP THREADPRIVATE(WORK) END MODULE INC_MODULE_GOOD3 SUBROUTINE SUB1(N) USE INC_MODULE_GOOD3 !$OMP PARALLEL PRIVATE(THE_SUM) ALLOCATE(WORK(N)) 248 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 CALL SUB2(THE_SUM) WRITE(*,*)THE_SUM !$OMP END PARALLEL END SUBROUTINE SUB1 SUBROUTINE SUB2(THE_SUM) USE INC_MODULE_GOOD3 WORK(:) = 10 THE_SUM=SUM(WORK) END SUBROUTINE SUB2 PROGRAM INC_GOOD3 N = 10 CALL SUB1(N) END PROGRAM INC_GOOD3 Fortran C/C++ 16 17 18 The following example illustrates initialization of threadprivate variables for class-type T. t1 is default constructed, t2 is constructed taking a constructor accepting one argument of integer type, t3 is copy constructed with argument f(): 19 Example A.27.4c 20 21 22 23 24 25 26 static T t1; #pragma omp threadprivate(t1) static T t2( 23 ); #pragma omp threadprivate(t2) static T t3 = f(); #pragma omp threadprivate(t3) 27 28 29 The following example illustrates the use of threadprivate for static class members. The threadprivate directive for a static class member must be placed inside the class definition. 30 Example A.27.5c 31 32 33 34 35 36 class T { public: static int i; #pragma omp threadprivate(i) }; C/C++ Appendix A Examples 249 1 C/C++ 2 A.28 Parallel Random Access Iterator Loop The following example shows a parallel random access iterator loop. 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Example A.28.1c #include <vector> void iterator_example() { std::vector<int> vec(23); std::vector<int>::iterator it; #pragma omp parallel for default(none) shared(vec) for (it = vec.begin(); it < vec.end(); it++) { // do work with *it // } } C/C++ 250 OpenMP API • Version 3.1 July 2011 1 Fortran 3 Fortran Restrictions on shared and private Clauses with Common Blocks 4 5 6 7 When a named common block is specified in a private, firstprivate, or lastprivate clause of a construct, none of its members may be declared in another data-sharing attribute clause on that construct. The following examples illustrate this point. For more information, see Section 2.9.3 on page 92. 8 The following example is conforming: 9 Example A.29.1f 2 10 11 12 13 14 15 16 17 18 19 20 21 A.29 SUBROUTINE COMMON_GOOD() COMMON /C/ X,Y REAL X, Y !$OMP !$OMP PARALLEL PRIVATE (/C/) ! do work here END PARALLEL !$OMP PARALLEL SHARED (X,Y) ! do work here !$OMP END PARALLEL END SUBROUTINE COMMON_GOOD 22 23 The following example is also conforming: 24 Example A.29.2f 25 26 27 28 29 30 31 32 33 34 35 36 37 SUBROUTINE COMMON_GOOD2() COMMON /C/ X,Y REAL X, Y INTEGER I !$OMP !$OMP !$OMP ! PARALLEL DO PRIVATE(/C/) DO I=1,1000 ! do work here ENDDO END DO Appendix A Examples 251 Fortran (cont.) 1 2 3 4 5 6 7 8 !$OMP 9 The following example is conforming: DO PRIVATE(X) DO I=1,1000 ! do work here ENDDO !$OMP END DO !$OMP END PARALLEL END SUBROUTINE COMMON_GOOD2 Example A.29.3f 10 11 12 13 14 15 16 17 18 19 20 21 SUBROUTINE COMMON_GOOD3() COMMON /C/ X,Y !$OMP !$OMP PARALLEL PRIVATE (/C/) ! do work here END PARALLEL !$OMP PARALLEL SHARED (/C/) ! do work here !$OMP END PARALLEL END SUBROUTINE COMMON_GOOD3 22 23 The following example is non-conforming because x is a constituent element of c: 24 Example A.29.4f 25 26 27 28 29 30 31 SUBROUTINE COMMON_WRONG() COMMON /C/ X,Y ! Incorrect because X is a constituent element of C !$OMP PARALLEL PRIVATE(/C/), SHARED(X) ! do work here !$OMP END PARALLEL END SUBROUTINE COMMON_WRONG 32 33 34 The following example is non-conforming because a common block may not be declared both shared and private: 35 Example A.29.5f 36 37 SUBROUTINE COMMON_WRONG2() COMMON /C/ X,Y 252 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 ! Incorrect: common block C cannot be declared both ! shared and private !$OMP PARALLEL PRIVATE (/C/), SHARED(/C/) ! do work here !$OMP END PARALLEL END SUBROUTINE COMMON_WRONG2 Fortran 8 A.30 The default(none) Clause 9 10 11 The following example distinguishes the variables that are affected by the default(none) clause from those that are not. For more information on the default clause, see Section 2.9.3.1 on page 93. 12 Example A.30.1c 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 #include <omp.h> int x, y, z[1000]; #pragma omp threadprivate(x) C/C++ void default_none(int a) { const int c = 1; int i = 0; #pragma omp parallel default(none) private(a) shared(z) { int j = omp_get_num_threads(); /* O.K. - j is declared within parallel region */ a = z[j]; /* O.K. - a is listed in private clause */ /* - z is listed in shared clause */ x = c; /* O.K. - x is threadprivate */ /* - c has const-qualified type */ z[i] = y; /* Error - cannot reference i or y here */ #pragma omp for firstprivate(y) /* Error - Cannot reference y in the firstprivate clause */ for (i=0; i<10 ; i++) { z[i] = i; /* O.K. - i is the loop iteration variable */ } z[i] = y; /* Error - cannot reference i or y here */ } } C/C++ Appendix A Examples 253 Fortran Example A.30.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 SUBROUTINE DEFAULT_NONE(A) INCLUDE "omp_lib.h" ! or USE OMP_LIB INTEGER A INTEGER X, Y, Z(1000) COMMON/BLOCKX/X COMMON/BLOCKY/Y COMMON/BLOCKZ/Z !$OMP THREADPRIVATE(/BLOCKX/) INTEGER I, J i = 1 !$OMP PARALLEL DEFAULT(NONE) PRIVATE(A) SHARED(Z) PRIVATE(J) J = OMP_GET_NUM_THREADS(); ! O.K. - J is listed in PRIVATE clause A = Z(J) ! O.K. - A is listed in PRIVATE clause ! - Z is listed in SHARED clause X = 1 ! O.K. - X is THREADPRIVATE Z(I) = Y ! Error - cannot reference I or Y here !$OMP DO firstprivate(y) ! Error - Cannot reference y in the firstprivate clause DO I = 1,10 Z(I) = I ! O.K. - I is the loop iteration variable END DO !$OMP Z(I) = Y ! Error - cannot reference I or Y here END PARALLEL END SUBROUTINE DEFAULT_NONE Fortran 254 OpenMP API • Version 3.1 July 2011 1 Fortran 2 3 A.31 Race Conditions Caused by Implied Copies of Shared Variables in Fortran 4 5 6 7 8 9 10 The following example contains a race condition, because the shared variable, which is an array section, is passed as an actual argument to a routine that has an assumed-size array as its dummy argument (see Section 2.9.3.2 on page 94). The subroutine call passing an array section argument may cause the compiler to copy the argument into a temporary location prior to the call and copy from the temporary location into the original variable when the subroutine returns. This copying would cause races in the parallel region. 11 Example A.31.1f 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 SUBROUTINE SHARED_RACE INCLUDE "omp_lib.h" ! or USE OMP_LIB REAL A(20) INTEGER MYTHREAD !$OMP PARALLEL SHARED(A) PRIVATE(MYTHREAD) MYTHREAD = OMP_GET_THREAD_NUM() IF (MYTHREAD .EQ. 0) THEN CALL SUB(A(1:10)) ! compiler may introduce writes to A(6:10) ELSE A(6:10) = 12 ENDIF !$OMP END PARALLEL END SUBROUTINE SHARED_RACE SUBROUTINE SUB(X) REAL X(*) X(1:5) = 4 END SUBROUTINE SUB Fortran Appendix A Examples 255 1 A.32 The private Clause 2 3 4 5 In the following example, the values of original list items i and j are retained on exit from the parallel region, while the private list items i and j are modified within the parallel construct. For more information on the private clause, see Section 2.9.3.3 on page 96. 6 Example A.32.1c 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 C/C++ #include <stdio.h> #include <assert.h> int main() { int i, j; int *ptr_i, *ptr_j; i = 1; j = 2; ptr_i = &i; ptr_j = &j; #pragma omp parallel private(i) firstprivate(j) { i = 3; j = j + 2; assert (*ptr_i == 1 && *ptr_j == 2); } assert(i == 1 && j == 2); return 0; } C/C++ Fortran Example A.32.1f 32 33 34 35 36 37 PROGRAM PRIV_EXAMPLE INTEGER I, J I = 1 J = 2 256 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 !$OMP !$OMP PARALLEL PRIVATE(I) FIRSTPRIVATE(J) I = 3 J = J + 2 END PARALLEL PRINT *, I, J ! I .eq. 1 .and. J .eq. 2 END PROGRAM PRIV_EXAMPLE Fortran 9 10 11 In the following example, all uses of the variable a within the loop construct in the routine f refer to a private list item a, while it is unspecified whether references to a in the routine g are to a private list item or the original list item. 12 Example A.32.2c 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 int a; C/C++ void g(int k) { a = k; /* Accessed in the region but outside of the construct; * therefore unspecified whether original or private list * item is modified. */ } void f(int n) { int a = 0; #pragma omp parallel for private(a) for (int i=1; i<n; i++) { a = i; g(a*2); /* Private copy of "a" */ } } C/C++ Appendix A Examples 257 Fortran Example A.32.2f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 MODULE PRIV_EXAMPLE2 REAL A CONTAINS SUBROUTINE G(K) REAL K A = K ! Accessed in the region but outside of the ! construct; therefore unspecified whether ! original or private list item is modified. END SUBROUTINE G SUBROUTINE F(N) INTEGER N REAL A !$OMP !$OMP INTEGER I PARALLEL DO PRIVATE(A) DO I = 1,N A = I CALL G(A*2) ENDDO END PARALLEL DO END SUBROUTINE F END MODULE PRIV_EXAMPLE2 Fortran The following example demonstrates that a list item that appears in a private clause in a parallel construct may also appear in a private clause in an enclosed worksharing construct, which results in an additional private copy. 28 29 30 258 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Example A.32.3c C/C++ #include <assert.h> void priv_example3() { int i, a; #pragma omp parallel private(a) { a = 1; #pragma omp parallel for private(a) for (i=0; i<10; i++) { a = 2; } assert(a == 1); } } C/C++ Fortran 18 19 20 21 22 23 24 25 26 27 28 29 30 31 Example A.32.3f SUBROUTINE PRIV_EXAMPLE3() INTEGER I, A !$OMP PARALLEL PRIVATE(A) A = 1 !$OMP PARALLEL DO PRIVATE(A) DO I = 1, 10 A = 2 END DO !$OMP END PARALLEL DO PRINT *, A ! Outer A still has value 1 !$OMP END PARALLEL END SUBROUTINE PRIV_EXAMPLE3 Fortran Appendix A Examples 259 1 Fortran 3 Fortran Restrictions on Storage Association with the private Clause 4 5 The following non-conforming examples illustrate the implications of the private clause rules with regard to storage association (see Section 2.9.3.3 on page 96). 6 Example A.33.1f 2 A.33 7 8 9 10 11 12 13 14 15 16 17 18 19 PROGRAM PRIV_RESTRICT COMMON /BLOCK/ X X = 1.0 !$OMP PARALLEL PRIVATE (X) X = 2.0 CALL SUB() !$OMP END PARALLEL END PROGRAM PRIV_RESTRICT 20 Example A.33.2f SUBROUTINE SUB() COMMON /BLOCK/ X PRINT *,X END SUBROUTINE SUB 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 ! X is undefined PROGRAM PRIV_RESTRICT2 COMMON /BLOCK2/ X X = 1.0 !$OMP !$OMP PARALLEL PRIVATE (X) X = 2.0 CALL SUB() END PARALLEL CONTAINS SUBROUTINE SUB() COMMON /BLOCK2/ Y PRINT *,X PRINT *,Y END SUBROUTINE SUB END PROGRAM PRIV_RESTRICT2 260 OpenMP API • Version 3.1 July 2011 ! X is undefined ! Y is undefined Fortran (cont.) 1 2 3 4 5 6 7 8 9 10 11 Example A.33.3f PROGRAM PRIV_RESTRICT3 EQUIVALENCE (X,Y) X = 1.0 !$OMP PARALLEL PRIVATE(X) PRINT *,Y Y = 10 PRINT *,X !$OMP END PARALLEL END PROGRAM PRIV_RESTRICT3 ! Y is undefined ! X is undefined 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 Example A.33.4f PROGRAM PRIV_RESTRICT4 INTEGER I, J INTEGER A(100), B(100) EQUIVALENCE (A(51), B(1)) !$OMP PARALLEL DO DEFAULT(PRIVATE) PRIVATE(I,J) LASTPRIVATE(A) DO I=1,100 DO J=1,100 B(J) = J - 1 ENDDO DO J=1,100 A(J) = J ENDDO ! B becomes undefined at this point DO J=1,50 B(J) = B(J) + 1 ! B is undefined ! A becomes undefined at this point ENDDO ENDDO !$OMP END PARALLEL DO ! The LASTPRIVATE write for A has ! undefined results PRINT *, B ! B is undefined since the LASTPRIVATE ! write of A was not defined END PROGRAM PRIV_RESTRICT4 Appendix A Examples 261 Example A.33.5f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 SUBROUTINE SUB1(X) DIMENSION X(10) ! ! ! ! ! This use of X does not conform to the specification. It would be legal Fortran 90, but the OpenMP private directive allows the compiler to break the sequence association that A had with the rest of the common block. FORALL (I = 1:10) X(I) = I END SUBROUTINE SUB1 PROGRAM PRIV_RESTRICT5 COMMON /BLOCK5/ A DIMENSION B(10) EQUIVALENCE (A,B(1)) ! the common block has to be at least 10 words A = 0 !$OMP PARALLEL PRIVATE(/BLOCK5/) ! ! ! ! ! !$OMP !$OMP !$OMP Without the private clause, we would be passing a member of a sequence that is at least ten elements long. With the private clause, A may no longer be sequence-associated. CALL SUB1(A) MASTER PRINT *, A END MASTER END PARALLEL END PROGRAM PRIV_RESTRICT5 Fortran 39 40 41 262 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 C/C++ A.34 C/C++ Arrays in a firstprivate Clause The following example illustrates the size and value of list items of array or pointer type in a firstprivate clause (Section 2.9.3.4 on page 98). The size of new list items is based on the type of the corresponding original list item, as determined by the base language. In this example: • The type of A is array of two arrays of two ints. • The type of B is adjusted to pointer to array of n ints, because it is a function parameter. • The type of C is adjusted to pointer to int, because it is a function parameter. • The type of D is array of two arrays of two ints. • The type of E is array of n arrays of n ints. 13 Note that B and E involve variable length array types. 14 15 16 The new items of array type are initialized as if each integer element of the original array is assigned to the corresponding element of the new array. Those of pointer type are initialized as if by assignment from the original item to the new item. 17 Appendix A Examples 263 Example A.34.1c 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 #include <assert.h> int A[2][2] = {1, 2, 3, 4}; void f(int n, int B[n][n], int C[]) { int D[2][2] = {1, 2, 3, 4}; int E[n][n]; assert(n >= 2); E[1][1] = 4; #pragma omp parallel firstprivate(B, C, D, E) { assert(sizeof(B) == sizeof(int (*)[n])); assert(sizeof(C) == sizeof(int*)); assert(sizeof(D) == 4 * sizeof(int)); assert(sizeof(E) == n * n * sizeof(int)); /* Private B and C have values of original B and C. */ assert(&B[1][1] == &A[1][1]); assert(&C[3] == &A[1][1]); assert(D[1][1] == 4); assert(E[1][1] == 4); } } int main() { f(2, A, A[0]); return 0; } C/C++ 33 A.35 The lastprivate Clause Correct execution sometimes depends on the value that the last iteration of a loop assigns to a variable. Such programs must list all such variables in a lastprivate clause (Section 2.9.3.5 on page 101) so that the values of the variables are the same as when the loop is executed sequentially. 34 35 36 37 264 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 C/C++ Example A.35.1c void lastpriv (int n, float *a, float *b) { int i; #pragma omp parallel { #pragma omp for lastprivate(i) for (i=0; i<n-1; i++) a[i] = b[i] + b[i+1]; } a[i]=b[i]; /* i == n-1 here */ } C/C++ Fortran 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 Example A.35.1f SUBROUTINE LASTPRIV(N, A, B) INTEGER N REAL A(*), B(*) INTEGER I !$OMP PARALLEL !$OMP DO LASTPRIVATE(I) DO I=1,N-1 A(I) = B(I) + B(I+1) ENDDO !$OMP END PARALLEL A(I) = B(I) ! I has the value of N here END SUBROUTINE LASTPRIV Fortran Appendix A Examples 265 1 A.36 The reduction Clause 2 3 4 The following example demonstrates the reduction clause (Section 2.9.3.6 on page 103); note that some reductions can be expressed in the loop in several ways, as shown for the max and min reductions below: 5 Example A.36.1c 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 C/C++ #include <math.h> void reduction1(float *x, int *y, int n) { int i, b, c; float a, d; a = 0.0; b = 0; c = y[0]; d = x[0]; #pragma omp parallel for private(i) shared(x, y, n) \ reduction(+:a) reduction(^:b) \ reduction(min:c) reduction(max:d) for (i=0; i<n; i++) { a += x[i]; b ^= y[i]; if (c > y[i]) c = y[i]; d = fmaxf(d,x[i]); } } C/C++ Fortran 25 Example A.36.1f 26 27 28 29 30 31 32 33 34 35 36 37 38 SUBROUTINE REDUCTION1(A, B, C, D, X, Y, N) REAL :: X(*), A, D INTEGER :: Y(*), N, B, C INTEGER :: I A = 0 B = 0 C = Y(1) D = X(1) !$OMP PARALLEL DO PRIVATE(I) SHARED(X, Y, N) REDUCTION(+:A) & !$OMP& REDUCTION(IEOR:B) REDUCTION(MIN:C) REDUCTION(MAX:D) DO I=1,N A = A + X(I) B = IEOR(B, Y(I)) 266 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 C = MIN(C, Y(I)) IF (D < X(I)) D = X(I) END DO END SUBROUTINE REDUCTION1 Fortran 6 7 A common implementation of the preceding example is to treat it as if it had been written as follows: 8 Example A.36.2c 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 C/C++ #include <limits.h> #include <math.h> void reduction2(float *x, int *y, int n) { int i, b, b_p, c, c_p; float a, a_p, d, d_p; a = 0.0f; b = 0; c = y[0]; d = x[0]; #pragma omp parallel shared(a, b, c, d, x, y, n) \ private(a_p, b_p, c_p, d_p) { a_p = 0.0f; b_p = 0; c_p = INT_MAX; d_p = -HUGE_VALF; #pragma omp for private(i) for (i=0; i<n; i++) { a_p += x[i]; b_p ^= y[i]; if (c_p > y[i]) c_p = y[i]; d_p = fmaxf(d_p,x[i]); } #pragma omp critical { a += a_p; b ^= b_p; if( c > c_p ) c = c_p; d = fmaxf(d,d_p); } } } C/C++ Appendix A Examples 267 Fortran Example A.36.2f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 SUBROUTINE REDUCTION2(A, B, C, D, X, Y, N) REAL :: X(*), A, D INTEGER :: Y(*), N, B, C REAL :: A_P, D_P INTEGER :: I, B_P, C_P A = 0 B = 0 C = Y(1) D = X(1) !$OMP PARALLEL SHARED(X, Y, A, B, C, D, N) & !$OMP& PRIVATE(A_P, B_P, C_P, D_P) A_P = 0.0 B_P = 0 C_P = HUGE(C_P) D_P = -HUGE(D_P) !$OMP DO PRIVATE(I) DO I=1,N A_P = A_P + X(I) B_P = IEOR(B_P, Y(I)) C_P = MIN(C_P, Y(I)) IF (D_P < X(I)) D_P = X(I) END DO !$OMP CRITICAL A = A + A_P B = IEOR(B, B_P) C = MIN(C, C_P) D = MAX(D, D_P) !$OMP END CRITICAL !$OMP END PARALLEL END SUBROUTINE REDUCTION2 32 33 34 The following program is non-conforming because the reduction is on the intrinsic procedure name MAX but that name has been redefined to be the variable named MAX. 35 Example A.36.3f 36 37 38 39 40 41 42 43 44 45 PROGRAM REDUCTION_WRONG MAX = HUGE(0) M = 0 !$OMP PARALLEL DO REDUCTION(MAX: M) ! MAX is no longer the intrinsic so this is non-conforming DO I = 1, 100 CALL SUB(M,I) END DO 268 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 END PROGRAM REDUCTION_WRONG SUBROUTINE SUB(M,I) M = MAX(M,I) END SUBROUTINE SUB 7 8 The following conforming program performs the reduction using the intrinsic procedure name MAX even though the intrinsic MAX has been renamed to REN. 9 Example A.36.4f 10 11 12 13 14 15 16 17 18 19 20 21 22 MODULE M INTRINSIC MAX END MODULE M 23 24 The following conforming program performs the reduction using intrinsic procedure name MAX even though the intrinsic MAX has been renamed to MIN. 25 Example A.36.5f 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 MODULE MOD INTRINSIC MAX, MIN END MODULE MOD PROGRAM REDUCTION3 USE M, REN => MAX N = 0 !$OMP PARALLEL DO REDUCTION(REN: N) DO I = 1, 100 N = MAX(N,I) END DO END PROGRAM REDUCTION3 ! still does MAX PROGRAM REDUCTION4 USE MOD, MIN=>MAX, MAX=>MIN REAL :: R R = -HUGE(0.0) !$OMP PARALLEL DO REDUCTION(MIN: R) DO I = 1, 1000 R = MIN(R, SIN(REAL(I))) END DO PRINT *, R END PROGRAM REDUCTION4 ! still does MAX Fortran Appendix A Examples 269 The following example is non-conforming because the initialization (a = 0) of the original list item a is not synchronized with the update of a as a result of the reduction computation in the for loop. Therefore, the example may print an incorrect value for a. 1 2 3 4 To avoid this problem, the initialization of the original list item a should complete before any update of a as a result of the reduction clause. This can be achieved by adding an explicit barrier after the assignment a = 0, or by enclosing the assignment a = 0 in a single directive (which has an implied barrier), or by initializing a before the start of the parallel region. 5 6 7 8 9 10 Example A.36.3c 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 #include <stdio.h> C/C++ int main (void) { int a, i; #pragma omp parallel shared(a) private(i) { #pragma omp master a = 0; // To avoid race conditions, add a barrier here. #pragma omp for reduction(+:a) for (i = 0; i < 10; i++) { a += i; } #pragma omp single printf ("Sum is %d\n", a); } } C/C++ 270 OpenMP API • Version 3.1 July 2011 Fortran Example A.36.6f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 INTEGER A, I !$OMP PARALLEL SHARED(A) PRIVATE(I) !$OMP MASTER A = 0 !$OMP END MASTER ! To avoid race conditions, add a barrier here. !$OMP DO REDUCTION(+:A) DO I= 0, 9 A = A + I END DO !$OMP SINGLE PRINT *, "Sum is ", A !$OMP END SINGLE !$OMP END PARALLEL END Fortran 23 24 25 26 A.37 The copyin Clause The copyin clause (see Section 2.9.4.1 on page 107) is used to initialize threadprivate data upon entry to a parallel region. The value of the threadprivate variable in the master thread is copied to the threadprivate variable of each other team member. Appendix A Examples 271 Example A.37.1c 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 C/C++ #include <stdlib.h> float* work; int size; float tol; #pragma omp threadprivate(work,size,tol) void build() { int i; work = (float*)malloc( sizeof(float)*size ); for( i = 0; i < size; ++i ) work[i] = tol; } void copyin_example( float t, int n ) { tol = t; size = n; #pragma omp parallel copyin(tol,size) { build(); } } C/C++ 272 OpenMP API • Version 3.1 July 2011 Fortran Example A.37.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 MODULE M REAL, POINTER, SAVE :: WORK(:) INTEGER :: SIZE REAL :: TOL !$OMP THREADPRIVATE(WORK,SIZE,TOL) END MODULE M SUBROUTINE COPYIN_EXAMPLE( T, N ) USE M REAL :: T INTEGER :: N TOL = T SIZE = N !$OMP PARALLEL COPYIN(TOL,SIZE) CALL BUILD !$OMP END PARALLEL END SUBROUTINE COPYIN_EXAMPLE SUBROUTINE BUILD USE M ALLOCATE(WORK(SIZE)) WORK = TOL END SUBROUTINE BUILD Fortran 25 A.38 The copyprivate Clause 26 27 28 29 30 31 The copyprivate clause (see Section 2.9.4.2 on page 109) can be used to broadcast values acquired by a single thread directly to all instances of the private variables in the other threads. In this example, if the routine is called from the sequential part, its behavior is not affected by the presence of the directives. If it is called from a parallel region, then the actual arguments with which a and b are associated must be private. 32 33 34 35 36 The thread that executes the structured block associated with the single construct broadcasts the values of the private variables a, b, x, and y from its implicit task's data environment to the data environments of the other implicit tasks in the thread team. The broadcast completes before any of the threads have left the barrier at the end of the construct. Appendix A Examples 273 Example A.38.1c 1 2 3 4 5 6 7 8 9 10 11 C/C++ #include <stdio.h> float x, y; #pragma omp threadprivate(x, y) void init(float a, float b ) { #pragma omp single copyprivate(a,b,x,y) { scanf("%f %f %f %f", &a, &b, &x, &y); } } C/C++ Fortran 12 Example A.38.1f 13 14 15 16 17 18 19 20 21 22 SUBROUTINE INIT(A,B) REAL A, B COMMON /XY/ X,Y !$OMP THREADPRIVATE (/XY/) !$OMP !$OMP SINGLE READ (11) A,B,X,Y END SINGLE COPYPRIVATE (A,B,/XY/) END SUBROUTINE INIT Fortran 274 OpenMP API • Version 3.1 July 2011 1 2 3 4 In this example, assume that the input must be performed by the master thread. Since the master construct does not support the copyprivate clause, it cannot broadcast the input value that is read. However, copyprivate is used to broadcast an address where the input value is stored. 5 Example A.38.2c 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 C/C++ #include <stdio.h> #include <stdlib.h> float read_next( ) { float * tmp; float return_val; #pragma omp single copyprivate(tmp) { tmp = (float *) malloc(sizeof(float)); } /* copies the pointer only */ #pragma omp master { scanf("%f", tmp); } #pragma omp barrier return_val = *tmp; #pragma omp barrier #pragma omp single nowait { free(tmp); } return return_val; } C/C++ Appendix A Examples 275 Fortran Example A.38.2f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 REAL FUNCTION READ_NEXT() REAL, POINTER :: TMP !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP SINGLE ALLOCATE (TMP) END SINGLE COPYPRIVATE (TMP) ! copies the pointer only MASTER READ (11) TMP END MASTER BARRIER READ_NEXT = TMP BARRIER SINGLE DEALLOCATE (TMP) END SINGLE NOWAIT END FUNCTION READ_NEXT Fortran 21 22 23 Suppose that the number of lock variables required within a parallel region cannot easily be determined prior to entering it. The copyprivate clause can be used to provide access to shared lock variables that are allocated within that parallel region. 24 Example A.38.3c 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 #include <stdio.h> #include <stdlib.h> #include <omp.h> C/C++ omp_lock_t *new_lock() { omp_lock_t *lock_ptr; #pragma omp single copyprivate(lock_ptr) { lock_ptr = (omp_lock_t *) malloc(sizeof(omp_lock_t)); omp_init_lock( lock_ptr ); } return lock_ptr; } C/C++ 276 OpenMP API • Version 3.1 July 2011 Fortran 1 Example A.38.3f 2 3 4 5 6 7 8 9 10 11 !$OMP 12 13 14 15 16 Note that the effect of the copyprivate clause on a variable with the allocatable attribute is different than on a variable with the pointer attribute. The value of A is copied (as if by intrinsic assignment) and the pointer B is copied (as if by pointer assignment) to the corresponding list items in the other implicit tasks belonging to the parallel region. 17 Example A.38.4f 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 FUNCTION NEW_LOCK() USE OMP_LIB ! or INCLUDE "omp_lib.h" INTEGER(OMP_LOCK_KIND), POINTER :: NEW_LOCK SINGLE ALLOCATE(NEW_LOCK) CALL OMP_INIT_LOCK(NEW_LOCK) !$OMP END SINGLE COPYPRIVATE(NEW_LOCK) END FUNCTION NEW_LOCK SUBROUTINE S(N) INTEGER N REAL, DIMENSION(:), ALLOCATABLE :: A REAL, DIMENSION(:), POINTER :: B !$OMP !$OMP ALLOCATE (A(N)) SINGLE ALLOCATE (B(N)) READ (11) A,B END SINGLE COPYPRIVATE(A,B) ! Variable A is private and is ! assigned the same value in each thread ! Variable B is shared !$OMP !$OMP BARRIER SINGLE DEALLOCATE (B) !$OMP END SINGLE NOWAIT END SUBROUTINE S Fortran Appendix A Examples 277 1 A.39 Nested Loop Constructs 2 3 4 The following example of loop construct nesting (see Section 2.10 on page 111) is conforming because the inner and outer loop regions bind to different parallel regions: 5 Example A.39.1c 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 C/C++ void work(int i, int j) {} void good_nesting(int n) { int i, j; #pragma omp parallel default(shared) { #pragma omp for for (i=0; i<n; i++) { #pragma omp parallel shared(i, n) { #pragma omp for for (j=0; j < n; j++) work(i, j); } } } } C/C++ 278 OpenMP API • Version 3.1 July 2011 Fortran 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 Example A.39.1f SUBROUTINE WORK(I, J) INTEGER I, J END SUBROUTINE WORK SUBROUTINE GOOD_NESTING(N) INTEGER N !$OMP !$OMP !$OMP !$OMP !$OMP !$OMP INTEGER I PARALLEL DEFAULT(SHARED) DO DO I = 1, N PARALLEL SHARED(I,N) DO DO J = 1, N CALL WORK(I,J) END DO END PARALLEL END DO END PARALLEL END SUBROUTINE GOOD_NESTING Fortran Appendix A Examples 279 1 The following variation of the preceding example is also conforming: 2 Example A.39.2c 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 C/C++ void work(int i, int j) {} void work1(int i, int n) { int j; #pragma omp parallel default(shared) { #pragma omp for for (j=0; j<n; j++) work(i, j); } } void good_nesting2(int n) { int i; #pragma omp parallel default(shared) { #pragma omp for for (i=0; i<n; i++) work1(i, n); } } C/C++ 280 OpenMP API • Version 3.1 July 2011 Fortran Example A.39.2f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 SUBROUTINE WORK(I, J) INTEGER I, J END SUBROUTINE WORK SUBROUTINE WORK1(I, N) INTEGER J !$OMP PARALLEL DEFAULT(SHARED) !$OMP DO DO J = 1, N CALL WORK(I,J) END DO !$OMP END PARALLEL END SUBROUTINE WORK1 SUBROUTINE GOOD_NESTING2(N) INTEGER N !$OMP PARALLEL DEFAULT(SHARED) !$OMP DO DO I = 1, N CALL WORK1(I, N) END DO !$OMP END PARALLEL END SUBROUTINE GOOD_NESTING2 Fortran 25 26 27 A.40 Restrictions on Nesting of Regions The examples in this section illustrate the region nesting rules. For more information on region nesting, see Section 2.10 on page 111. Appendix A Examples 281 1 2 The following example is non-conforming because the inner and outer loop regions are closely nested: 3 Example A.40.1c 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 C/C++ void work(int i, int j) {} void wrong1(int n) { #pragma omp parallel default(shared) { int i, j; #pragma omp for for (i=0; i<n; i++) { /* incorrect nesting of loop regions */ #pragma omp for for (j=0; j<n; j++) work(i, j); } } } C/C++ Fortran Example A.40.1f 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 SUBROUTINE WORK(I, J) INTEGER I, J END SUBROUTINE WORK SUBROUTINE WRONG1(N) INTEGER N INTEGER I,J PARALLEL DEFAULT(SHARED) DO DO I = 1, N !$OMP DO ! incorrect nesting of loop regions DO J = 1, N CALL WORK(I,J) END DO END DO !$OMP END PARALLEL END SUBROUTINE WRONG1 !$OMP !$OMP Fortran 282 OpenMP API • Version 3.1 July 2011 1 The following orphaned version of the preceding example is also non-conforming: 2 Example A.40.2c 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 C/C++ void work(int i, int j) {} void work1(int i, int n) { int j; /* incorrect nesting of loop regions */ #pragma omp for for (j=0; j<n; j++) work(i, j); } void wrong2(int n) { #pragma omp parallel default(shared) { int i; #pragma omp for for (i=0; i<n; i++) work1(i, n); } } C/C++ Fortran 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 Example A.40.2f !$OMP !$OMP !$OMP !$OMP SUBROUTINE WORK1(I,N) INTEGER I, N INTEGER J DO ! incorrect nesting of loop regions DO J = 1, N CALL WORK(I,J) END DO END SUBROUTINE WORK1 SUBROUTINE WRONG2(N) INTEGER N INTEGER I PARALLEL DEFAULT(SHARED) DO DO I = 1, N CALL WORK1(I,N) END DO END PARALLEL END SUBROUTINE WRONG2 Fortran Appendix A Examples 283 1 2 The following example is non-conforming because the loop and single regions are closely nested: 3 Example A.40.3c 4 5 6 7 8 9 10 11 12 13 14 15 16 17 C/C++ void work(int i, int j) {} void wrong3(int n) { #pragma omp parallel default(shared) { int i; #pragma omp for for (i=0; i<n; i++) { /* incorrect nesting of regions */ #pragma omp single work(i, 0); } } } C/C++ Fortran Example A.40.3f 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 SUBROUTINE WRONG3(N) INTEGER N INTEGER I PARALLEL DEFAULT(SHARED) DO DO I = 1, N !$OMP SINGLE ! incorrect nesting of regions CALL WORK(I, 1) !$OMP END SINGLE END DO !$OMP END PARALLEL END SUBROUTINE WRONG3 !$OMP !$OMP Fortran 284 OpenMP API • Version 3.1 July 2011 1 2 The following example is non-conforming because a barrier region cannot be closely nested inside a loop region: 3 Example A.40.4c 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 C/C++ void work(int i, int j) {} void wrong4(int n) { #pragma omp parallel default(shared) { int i; #pragma omp for for (i=0; i<n; i++) { work(i, 0); /* incorrect nesting of barrier region in a loop region */ #pragma omp barrier work(i, 1); } } } C/C++ Fortran 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 Example A.40.4f SUBROUTINE WRONG4(N) INTEGER N INTEGER I PARALLEL DEFAULT(SHARED) DO DO I = 1, N CALL WORK(I, 1) ! incorrect nesting of barrier region in a loop region !$OMP BARRIER CALL WORK(I, 2) END DO !$OMP END PARALLEL END SUBROUTINE WRONG4 !$OMP !$OMP Fortran Appendix A Examples 285 1 2 3 The following example is non-conforming because the barrier region cannot be closely nested inside the critical region. If this were permitted, it would result in deadlock due to the fact that only one thread at a time can enter the critical region: 4 Example A.40.5c 5 6 7 8 9 10 11 12 13 14 15 16 17 18 C/C++ void work(int i, int j) {} void wrong5(int n) { #pragma omp parallel { #pragma omp critical { work(n, 0); /* incorrect nesting of barrier region in a critical region */ #pragma omp barrier work(n, 1); } } } C/C++ Fortran Example A.40.5f 19 20 21 22 23 24 25 26 27 28 29 30 31 SUBROUTINE WRONG5(N) INTEGER N !$OMP !$OMP PARALLEL DEFAULT(SHARED) CRITICAL CALL WORK(N,1) ! incorrect nesting of barrier region in a critical region !$OMP BARRIER CALL WORK(N,2) !$OMP END CRITICAL !$OMP END PARALLEL END SUBROUTINE WRONG5 Fortran 286 OpenMP API • Version 3.1 July 2011 1 2 3 The following example is non-conforming because the barrier region cannot be closely nested inside the single region. If this were permitted, it would result in deadlock due to the fact that only one thread executes the single region: 4 Example A.40.6c 5 6 7 8 9 10 11 12 13 14 15 16 17 18 C/C++ void work(int i, int j) {} void wrong6(int n) { #pragma omp parallel { #pragma omp single { work(n, 0); /* incorrect nesting of barrier region in a single region */ #pragma omp barrier work(n, 1); } } } C/C++ Fortran 19 20 21 22 23 24 25 26 27 28 29 30 31 Example A.40.6f SUBROUTINE WRONG6(N) INTEGER N !$OMP !$OMP PARALLEL DEFAULT(SHARED) SINGLE CALL WORK(N,1) ! incorrect nesting of barrier region in a single region !$OMP BARRIER CALL WORK(N,2) !$OMP END SINGLE !$OMP END PARALLEL END SUBROUTINE WRONG6 Fortran Appendix A Examples 287 2 The omp_set_dynamic and omp_set_num_threads Routines 3 4 5 6 7 8 Some programs rely on a fixed, prespecified number of threads to execute correctly. Because the default setting for the dynamic adjustment of the number of threads is implementation defined, such programs can choose to turn off the dynamic threads capability and set the number of threads explicitly to ensure portability. The following example shows how to do this using omp_set_dynamic (Section 3.2.7 on page 123), and omp_set_num_threads (Section 3.2.1 on page 116). 9 10 11 12 13 14 15 In this example, the program executes correctly only if it is executed by 16 threads. If the implementation is not capable of supporting 16 threads, the behavior of this example is implementation defined (see Algorithm 2.1 on page 36). Note that the number of threads executing a parallel region remains constant during the region, regardless of the dynamic threads setting. The dynamic threads mechanism determines the number of threads to use at the start of the parallel region and keeps it constant for the duration of the region. 16 Example A.41.1c 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 #include <omp.h> #include <stdlib.h> 1 A.41 C/C++ void do_by_16(float *x, int iam, int ipoints) {} void dynthreads(float *x, int npoints) { int iam, ipoints; omp_set_dynamic(0); omp_set_num_threads(16); #pragma omp parallel shared(x, npoints) private(iam, ipoints) { if (omp_get_num_threads() != 16) abort(); iam = omp_get_thread_num(); ipoints = npoints/16; do_by_16(x, iam, ipoints); } } C/C++ 288 OpenMP API • Version 3.1 July 2011 Fortran Example A.41.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 SUBROUTINE DO_BY_16(X, IAM, IPOINTS) REAL X(*) INTEGER IAM, IPOINTS END SUBROUTINE DO_BY_16 SUBROUTINE DYNTHREADS(X, NPOINTS) INCLUDE "omp_lib.h" ! or USE OMP_LIB INTEGER NPOINTS REAL X(NPOINTS) INTEGER IAM, IPOINTS CALL OMP_SET_DYNAMIC(.FALSE.) CALL OMP_SET_NUM_THREADS(16) !$OMP PARALLEL SHARED(X,NPOINTS) PRIVATE(IAM, IPOINTS) IF (OMP_GET_NUM_THREADS() .NE. 16) THEN STOP ENDIF IAM = OMP_GET_THREAD_NUM() IPOINTS = NPOINTS/16 CALL DO_BY_16(X,IAM,IPOINTS) !$OMP END PARALLEL END SUBROUTINE DYNTHREADS Fortran 32 33 34 35 36 A.42 The omp_get_num_threads Routine In the following example, the omp_get_num_threads call (see Section 3.2.2 on page 117) returns 1 in the sequential part of the code, so np will always be equal to 1. To determine the number of threads that will be deployed for the parallel region, the call should be inside the parallel region. Appendix A Examples 289 1 Example A.42.1c 2 3 4 5 6 7 8 9 10 11 12 13 14 #include <omp.h> void work(int i); C/C++ void incorrect() { int np, i; np = omp_get_num_threads(); /* misplaced */ #pragma omp parallel for schedule(static) for (i=0; i < np; i++) work(i); } C/C++ Fortran Example A.42.1f 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 SUBROUTINE WORK(I) INTEGER I I = I + 1 END SUBROUTINE WORK SUBROUTINE INCORRECT() INCLUDE "omp_lib.h" INTEGER I, NP ! or USE OMP_LIB NP = OMP_GET_NUM_THREADS() !misplaced: will return 1 PARALLEL DO SCHEDULE(STATIC) DO I = 0, NP-1 CALL WORK(I) ENDDO !$OMP END PARALLEL DO END SUBROUTINE INCORRECT !$OMP Fortran 290 OpenMP API • Version 3.1 July 2011 1 2 The following example shows how to rewrite this program without including a query for the number of threads: 3 Example A.42.2c 4 5 6 7 8 9 10 11 12 13 14 15 16 #include <omp.h> void work(int i); C/C++ void correct() { int i; #pragma omp parallel private(i) { i = omp_get_thread_num(); work(i); } } C/C++ Fortran 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 Example A.42.2f SUBROUTINE WORK(I) INTEGER I I = I + 1 END SUBROUTINE WORK SUBROUTINE CORRECT() INCLUDE "omp_lib.h" INTEGER I !$OMP !$OMP ! or USE OMP_LIB PARALLEL PRIVATE(I) I = OMP_GET_THREAD_NUM() CALL WORK(I) END PARALLEL END SUBROUTINE CORRECT Fortran Appendix A Examples 291 1 A.43 The omp_init_lock Routine 2 3 The following example demonstrates how to initialize an array of locks in a parallel region by using omp_init_lock (Section 3.3.1 on page 143). 4 Example A.43.1c 5 6 7 8 9 10 11 12 13 14 15 16 17 18 C/C++ #include <omp.h> omp_lock_t *new_locks() { int i; omp_lock_t *lock = new omp_lock_t[1000]; #pragma omp parallel for private(i) for (i=0; i<1000; i++) { omp_init_lock(&lock[i]); } return lock; } C/C++ Fortran Example A.43.1f 19 20 21 22 23 24 25 26 27 28 29 30 31 32 FUNCTION NEW_LOCKS() USE OMP_LIB ! or INCLUDE "omp_lib.h" INTEGER(OMP_LOCK_KIND), DIMENSION(1000) :: NEW_LOCKS INTEGER I !$OMP !$OMP PARALLEL DO PRIVATE(I) DO I=1,1000 CALL OMP_INIT_LOCK(NEW_LOCKS(I)) END DO END PARALLEL DO END FUNCTION NEW_LOCKS Fortran 292 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 A.44 Ownership of Locks Ownership of locks has changed since OpenMP 2.5. In OpenMP 2.5, locks are owned by threads; so a lock released by the omp_unset_lock routine must be owned by the same thread executing the routine. With OpenMP 3.0, locks are owned by task regions; so a lock released by the omp_unset_lock routine in a task region must be owned by the same task region. 7 8 9 10 11 12 13 This change in ownership requires extra care when using locks. The following program is conforming in OpenMP 2.5 because the thread that releases the lock lck in the parallel region is the same thread that acquired the lock in the sequential part of the program (master thread of parallel region and the initial thread are the same). However, it is not conforming in OpenMP 3.0 and 3.1, because the task region that releases the lock lck is different from the task region that acquires the lock. 14 Example A.44.1c 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 #include <stdlib.h> #include <stdio.h> #include <omp.h> C/C++ int main() { int x; omp_lock_t lck; omp_init_lock (&lck); omp_set_lock (&lck); x = 0; #pragma omp parallel shared (x) { #pragma omp master { x = x + 1; omp_unset_lock (&lck); } /* Some more stuff. */ } omp_destroy_lock (&lck); return 0; } C/C++ Appendix A Examples 293 Fortran Example A.44.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 program lock use omp_lib integer :: x integer (kind=omp_lock_kind) :: lck call omp_init_lock (lck) call omp_set_lock(lck) x = 0 !$omp parallel shared (x) !$omp master x = x + 1 call omp_unset_lock(lck) !$omp end master ! Some more stuff. !$omp end parallel call omp_destroy_lock(lck) end Fortran 22 23 A.45 Simple Lock Routines In the following example (for Section 3.3 on page 141), the lock routines cause the threads to be idle while waiting for entry to the first critical section, but to do other work while waiting for entry to the second. The omp_set_lock function blocks, but the omp_test_lock function does not, allowing the work in skip to be done. 24 25 26 27 294 OpenMP API • Version 3.1 July 2011 C/C++ 1 2 Note that the argument to the lock routines should have type omp_lock_t, and that there is no need to flush it. 3 Example A.45.1c 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 #include <stdio.h> #include <omp.h> void skip(int i) {} void work(int i) {} int main() { omp_lock_t lck; int id; omp_init_lock(&lck); #pragma omp parallel shared(lck) private(id) { id = omp_get_thread_num(); omp_set_lock(&lck); /* only one thread at a time can execute this printf */ printf("My thread id is %d.\n", id); omp_unset_lock(&lck); while (! omp_test_lock(&lck)) { skip(id); /* we do not yet have the lock, so we must do something else */ } work(id); /* we now have the lock and can do the work */ omp_unset_lock(&lck); } omp_destroy_lock(&lck); return 0; } C/C++ Appendix A Examples 295 Fortran 1 Note that there is no need to flush the lock variable. 2 Example A.45.1f 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 SUBROUTINE SKIP(ID) END SUBROUTINE SKIP SUBROUTINE WORK(ID) END SUBROUTINE WORK PROGRAM SIMPLELOCK INCLUDE "omp_lib.h" ! or USE OMP_LIB INTEGER(OMP_LOCK_KIND) LCK INTEGER ID CALL OMP_INIT_LOCK(LCK) !$OMP PARALLEL SHARED(LCK) PRIVATE(ID) ID = OMP_GET_THREAD_NUM() CALL OMP_SET_LOCK(LCK) PRINT *, 'My thread id is ', ID CALL OMP_UNSET_LOCK(LCK) DO WHILE (.NOT. OMP_TEST_LOCK(LCK)) CALL SKIP(ID) ! We do not yet have the lock ! so we must do something else END DO CALL WORK(ID) ! We now have the lock ! and can do the work CALL OMP_UNSET_LOCK( LCK ) !$OMP END PARALLEL CALL OMP_DESTROY_LOCK( LCK ) END PROGRAM SIMPLELOCK Fortran 296 OpenMP API • Version 3.1 July 2011 1 A.46 Nestable Lock Routines 2 3 The following example (for Section 3.3 on page 141) demonstrates how a nestable lock can be used to synchronize updates both to a whole structure and to one of its members. 4 Example A.46.1c 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 C/C++ #include <omp.h> typedef struct { int a,b; omp_nest_lock_t lck; } pair; int work1(); int work2(); int work3(); void incr_a(pair *p, int a) { /* Called only from incr_pair, no need to lock. */ p->a += a; } void incr_b(pair *p, int b) { /* Called both from incr_pair and elsewhere, */ /* so need a nestable lock. */ omp_set_nest_lock(&p->lck); p->b += b; omp_unset_nest_lock(&p->lck); } void incr_pair(pair *p, int a, int b) { omp_set_nest_lock(&p->lck); incr_a(p, a); incr_b(p, b); omp_unset_nest_lock(&p->lck); } void nestlock(pair *p) { #pragma omp parallel sections { #pragma omp section incr_pair(p, work1(), work2()); #pragma omp section incr_b(p, work3()); } } C/C++ Appendix A Examples 297 Fortran Example A.46.1f 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 MODULE DATA USE OMP_LIB, ONLY: OMP_NEST_LOCK_KIND TYPE LOCKED_PAIR INTEGER A INTEGER B INTEGER (OMP_NEST_LOCK_KIND) LCK END TYPE END MODULE DATA SUBROUTINE INCR_A(P, A) ! called only from INCR_PAIR, no need to lock USE DATA TYPE(LOCKED_PAIR) :: P INTEGER A P%A = P%A + A END SUBROUTINE INCR_A SUBROUTINE INCR_B(P, B) ! called from both INCR_PAIR and elsewhere, ! so we need a nestable lock USE OMP_LIB ! or INCLUDE "omp_lib.h" USE DATA TYPE(LOCKED_PAIR) :: P INTEGER B CALL OMP_SET_NEST_LOCK(P%LCK) P%B = P%B + B CALL OMP_UNSET_NEST_LOCK(P%LCK) END SUBROUTINE INCR_B SUBROUTINE INCR_PAIR(P, A, B) USE OMP_LIB ! or INCLUDE "omp_lib.h" USE DATA TYPE(LOCKED_PAIR) :: P INTEGER A INTEGER B CALL OMP_SET_NEST_LOCK(P%LCK) CALL INCR_A(P, A) CALL INCR_B(P, B) CALL OMP_UNSET_NEST_LOCK(P%LCK) END SUBROUTINE INCR_PAIR SUBROUTINE NESTLOCK(P) USE OMP_LIB ! or INCLUDE "omp_lib.h" USE DATA TYPE(LOCKED_PAIR) :: P INTEGER WORK1, WORK2, WORK3 EXTERNAL WORK1, WORK2, WORK3 298 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 !$OMP PARALLEL SECTIONS !$OMP SECTION CALL INCR_PAIR(P, WORK1(), WORK2()) SECTION CALL INCR_B(P, WORK3()) END PARALLEL SECTIONS !$OMP !$OMP END SUBROUTINE NESTLOCK Fortran 10 Appendix A Examples 299 1 This page intentionally left blank. 2 300 OpenMP API • Version 3.1 July 2011 1 APPENDIX B 3 Stubs for Runtime Library Routines 4 5 6 7 8 This section provides stubs for the runtime library routines defined in the OpenMP API. The stubs are provided to enable portability to platforms that do not support the OpenMP API. On these platforms, OpenMP programs must be linked with a library containing these stub routines. The stub routines assume that the directives in the OpenMP program are ignored. As such, they emulate serial semantics. 9 10 11 Note that the lock variable that appears in the lock routines must be accessed exclusively through these routines. It should not be initialized or otherwise modified in the user program. 12 13 14 15 In an actual implementation the lock variable might be used to hold the address of an allocated memory block, but here it is used to hold an integer value. Users should not make assumptions about mechanisms used by OpenMP implementations to implement locks based on the scheme used by the stub procedures. 2 Fortran 16 17 18 19 Note – In order to be able to compile the Fortran stubs file, the include file omp_lib.h was split into two files: omp_lib_kinds.h and omp_lib.h and the omp_lib_kinds.h file included where needed. There is no requirement for the implementation to provide separate files. Fortran 301 1 B.1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 C/C++ Stub Routines #include <stdio.h> #include <stdlib.h> #include "omp.h" void omp_set_num_threads(int num_threads) { } int omp_get_num_threads(void) { return 1; } int omp_get_max_threads(void) { return 1; } int omp_get_thread_num(void) { return 0; } int omp_get_num_procs(void) { return 1; } int omp_in_parallel(void) { return 0; } void omp_set_dynamic(int dynamic_threads) { } int omp_get_dynamic(void) { return 0; } void omp_set_nested(int nested) { } 302 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 int omp_get_nested(void) { return 0; } void omp_set_schedule(omp_sched_t kind, int modifier) { } void omp_get_schedule(omp_sched_t *kind, int *modifier) { *kind = omp_sched_static; *modifier = 0; } int omp_get_thread_limit(void) { return 1; } void omp_set_max_active_levels(int max_active_levels) { } int omp_get_max_active_levels(void) { return 0; } int omp_get_level(void) { return 0; } int omp_get_ancestor_thread_num(int level) { if (level == 0) { return 0; } else { return -1; } } Appendix B Stubs for Runtime Library Routines 303 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 int omp_get_team_size(int level) { if (level == 0) { return 1; } else { return -1; } } int omp_get_active_level(void) { return 0; } int omp_in_final(void) { return 1; } struct __omp_lock { int lock; }; enum { UNLOCKED = -1, INIT, LOCKED }; void omp_init_lock(omp_lock_t *arg) { struct __omp_lock *lock = (struct __omp_lock *)arg; lock->lock = UNLOCKED; } void omp_destroy_lock(omp_lock_t *arg) { struct __omp_lock *lock = (struct __omp_lock *)arg; lock->lock = INIT; } 304 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 void omp_set_lock(omp_lock_t *arg) { struct __omp_lock *lock = (struct __omp_lock *)arg; if (lock->lock == UNLOCKED) { lock->lock = LOCKED; } else if (lock->lock == LOCKED) { fprintf(stderr, "error: deadlock in using lock variable\n"); exit(1); } else { fprintf(stderr, "error: lock not initialized\n"); exit(1); } } void omp_unset_lock(omp_lock_t *arg) { struct __omp_lock *lock = (struct __omp_lock *)arg; if (lock->lock == LOCKED) { lock->lock = UNLOCKED; } else if (lock->lock == UNLOCKED) { fprintf(stderr, "error: lock not set\n"); exit(1); } else { fprintf(stderr, "error: lock not initialized\n"); exit(1); } } Appendix B Stubs for Runtime Library Routines 305 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 int omp_test_lock(omp_lock_t *arg) { struct __omp_lock *lock = (struct __omp_lock *)arg; if (lock->lock == UNLOCKED) { lock->lock = LOCKED; return 1; } else if (lock->lock == LOCKED) { return 0; } else { fprintf(stderr, "error: lock not initialized\n"); exit(1); } } struct __omp_nest_lock { short owner; short count; }; enum { NOOWNER = -1, MASTER = 0 }; void omp_init_nest_lock(omp_nest_lock_t *arg) { struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg; nlock->owner = NOOWNER; nlock->count = 0; } void omp_destroy_nest_lock(omp_nest_lock_t *arg) { struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg; nlock->owner = NOOWNER; nlock->count = UNLOCKED; } 306 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 void omp_set_nest_lock(omp_nest_lock_t *arg) { struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg; if (nlock->owner == MASTER && nlock->count >= 1) { nlock->count++; } else if (nlock->owner == NOOWNER && nlock->count == 0) { nlock->owner = MASTER; nlock->count = 1; } else { fprintf(stderr, "error: lock corrupted or not initialized\n"); exit(1); } } void omp_unset_nest_lock(omp_nest_lock_t *arg) { struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg; if (nlock->owner == MASTER && nlock->count >= 1) { nlock->count--; if (nlock->count == 0) { nlock->owner = NOOWNER; } } else if (nlock->owner == NOOWNER && nlock->count == 0) { fprintf(stderr, "error: lock not set\n"); exit(1); } else { fprintf(stderr, "error: lock corrupted or not initialized\n"); exit(1); } } int omp_test_nest_lock(omp_nest_lock_t *arg) { struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg; omp_set_nest_lock(arg); return nlock->count; } Appendix B Stubs for Runtime Library Routines 307 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 double omp_get_wtime(void) { /* This function does not provide a working * wallclock timer. Replace it with a version * customized for the target machine. */ return 0.0; } double omp_get_wtick(void) { /* This function does not provide a working * clock tick function. Replace it with * a version customized for the target machine. */ return 365. * 86400.; } 18 308 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 B.2 Fortran Stub Routines C23456 subroutine omp_set_num_threads(num_threads) integer num_threads return end subroutine integer function omp_get_num_threads() omp_get_num_threads = 1 return end function integer function omp_get_max_threads() omp_get_max_threads = 1 return end function integer function omp_get_thread_num() omp_get_thread_num = 0 return end function integer function omp_get_num_procs() omp_get_num_procs = 1 return end function logical function omp_in_parallel() omp_in_parallel = .false. return end function subroutine omp_set_dynamic(dynamic_threads) logical dynamic_threads return end subroutine logical function omp_get_dynamic() omp_get_dynamic = .false. return end function subroutine omp_set_nested(nested) logical nested return end subroutine Appendix B Stubs for Runtime Library Routines 309 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 logical function omp_get_nested() omp_get_nested = .false. return end function subroutine omp_set_schedule(kind, modifier) include 'omp_lib_kinds.h' integer (kind=omp_sched_kind) kind integer modifier return end subroutine subroutine omp_get_schedule(kind, modifier) include 'omp_lib_kinds.h' integer (kind=omp_sched_kind) kind integer modifier kind = omp_sched_static modifier = 0 return end subroutine integer function omp_get_thread_limit() omp_get_thread_limit = 1 return end function subroutine omp_set_max_active_levels( level ) integer level end subroutine integer function omp_get_max_active_levels() omp_get_max_active_levels = 0 return end function integer function omp_get_level() omp_get_level = 0 return end function integer function omp_get_ancestor_thread_num( level ) integer level if ( level .eq. 0 ) then omp_get_ancestor_thread_num = 0 else omp_get_ancestor_thread_num = -1 end if return end function 310 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 integer function omp_get_team_size( level ) integer level if ( level .eq. 0 ) then omp_get_team_size = 1 else omp_get_team_size = -1 end if return end function integer function omp_get_active_level() omp_get_active_level = 0 return end function logical function omp_in_final() omp_in_final = .true. return end function subroutine omp_init_lock(lock) ! lock is 0 if the simple lock is not initialized ! -1 if the simple lock is initialized but not set ! 1 if the simple lock is set include 'omp_lib_kinds.h' integer(kind=omp_lock_kind) lock lock = -1 return end subroutine subroutine omp_destroy_lock(lock) include 'omp_lib_kinds.h' integer(kind=omp_lock_kind) lock lock = 0 return end subroutine subroutine omp_set_lock(lock) include 'omp_lib_kinds.h' integer(kind=omp_lock_kind) lock if (lock .eq. -1) then lock = 1 elseif (lock .eq. 1) then print *, 'error: deadlock in using lock variable' stop else print *, 'error: lock not initialized' stop endif return end subroutine Appendix B Stubs for Runtime Library Routines 311 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 subroutine omp_unset_lock(lock) include 'omp_lib_kinds.h' integer(kind=omp_lock_kind) lock if (lock .eq. 1) then lock = -1 elseif (lock .eq. -1) then print *, 'error: lock not set' stop else print *, 'error: lock not initialized' stop endif return end subroutine logical function omp_test_lock(lock) include 'omp_lib_kinds.h' integer(kind=omp_lock_kind) lock if (lock .eq. -1) then lock = 1 omp_test_lock = .true. elseif (lock .eq. 1) then omp_test_lock = .false. else print *, 'error: lock not initialized' stop endif return end function subroutine omp_init_nest_lock(nlock) ! nlock is ! 0 if the nestable lock is not initialized ! -1 if the nestable lock is initialized but not set ! 1 if the nestable lock is set ! no use count is maintained include 'omp_lib_kinds.h' integer(kind=omp_nest_lock_kind) nlock nlock = -1 return end subroutine 312 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 subroutine omp_destroy_nest_lock(nlock) include 'omp_lib_kinds.h' integer(kind=omp_nest_lock_kind) nlock nlock = 0 return end subroutine subroutine omp_set_nest_lock(nlock) include 'omp_lib_kinds.h' integer(kind=omp_nest_lock_kind) nlock if (nlock .eq. -1) nlock = 1 elseif (nlock .eq. print *, 'error: stop else print *, 'error: stop endif then 0) then nested lock not initialized' deadlock using nested lock variable' return end subroutine subroutine omp_unset_nest_lock(nlock) include 'omp_lib_kinds.h' integer(kind=omp_nest_lock_kind) nlock if (nlock .eq. 1) then nlock = -1 elseif (nlock .eq. 0) then print *, 'error: nested lock not initialized' stop else print *, 'error: nested lock not set' stop endif return end subroutine Appendix B Stubs for Runtime Library Routines 313 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 integer function omp_test_nest_lock(nlock) include 'omp_lib_kinds.h' integer(kind=omp_nest_lock_kind) nlock if (nlock .eq. -1) then nlock = 1 omp_test_nest_lock = 1 elseif (nlock .eq. 1) then omp_test_nest_lock = 0 else print *, 'error: nested lock not initialized' stop endif return end function double precision function omp_get_wtime() ! this function does not provide a working ! wall clock timer. replace it with a version ! customized for the target machine. omp_get_wtime = 0.0d0 return end function double precision function omp_get_wtick() ! this function does not provide a working ! clock tick function. replace it with ! a version customized for the target machine. double precision one_year parameter (one_year=365.d0*86400.d0) omp_get_wtick = one_year return end function 314 OpenMP API • Version 3.1 July 2011 1 APPENDIX C OpenMP C and C++ Grammar 2 3 4 C.1 Notation 5 6 The grammar rules consist of the name for a non-terminal, followed by a colon, followed by replacement alternatives on separate lines. 7 8 The syntactic expression termopt indicates that the term is optional within the replacement. 9 10 The syntactic expression termoptseq is equivalent to term-seqopt with the following additional rules: 11 term-seq : 12 term 13 term-seq term 14 term-seq , term 315 1 C.2 Rules The notation is described in Section 6.1 of the C standard. This grammar appendix shows the extensions to the base language grammar for the OpenMP C and C++ directives. 2 3 4 5 6 /* in C++ (ISO/IEC 14882:1998) */ 7 statement-seq: statement 8 openmp-directive 9 10 statement-seq statement 11 statement-seq openmp-directive 12 13 14 /* in C90 (ISO/IEC 9899:1990) */ 15 statement-list: statement 16 17 openmp-directive 18 statement-list statement 19 statement-list openmp-directive 20 21 22 /* in C99 (ISO/IEC 9899:1999) */ 23 block-item: declaration 24 25 statement 26 openmp-directive 27 316 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 statement: /* standard statements */ openmp-construct openmp-construct: parallel-construct 7 for-construct 8 sections-construct 9 single-construct 10 parallel-for-construct 11 parallel-sections-construct 12 task-construct 13 master-construct 14 critical-construct 15 atomic-construct 16 ordered-construct 17 openmp-directive: 18 barrier-directive 19 taskwait-directive 20 taskyield-directive 21 flush-directive 22 structured-block: 23 statement 24 25 26 27 parallel-construct: parallel-directive structured-block parallel-directive: # pragma omp parallel parallel-clauseoptseq new-line 28 Appendix C OpenMP C and C++ Grammar 317 parallel-clause: 1 unique-parallel-clause 2 3 data-default-clause 4 data-privatization-clause 5 data-privatization-in-clause 6 data-sharing-clause 7 data-reduction-clause unique-parallel-clause: 8 if ( expression ) 9 10 num_threads ( expression ) 11 copyin ( variable-list ) for-construct: 12 for-directive iteration-statement 13 for-directive: 14 # pragma omp for for-clauseoptseq new-line 15 for-clause: 16 unique-for-clause 17 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 ) 28 318 OpenMP API • Version 3.1 July 2011 1 schedule-kind: 2 static 3 dynamic 4 guided 5 auto 6 runtime 7 8 9 10 11 sections-construct: sections-directive section-scope sections-directive: # pragma omp sections sections-clauseoptseq new-line sections-clause: 12 data-privatization-clause 13 data-privatization-in-clause 14 data-privatization-out-clause 15 data-reduction-clause 16 nowait 17 section-scope: 18 19 { section-sequence } section-sequence: 20 section-directiveopt structured-block 21 section-sequence section-directive structured-block 22 23 24 25 26 27 section-directive: # pragma omp section new-line single-construct: single-directive structured-block single-directive: # pragma omp single single-clauseoptseq new-line 28 Appendix C OpenMP C and C++ Grammar 319 single-clause: 1 unique-single-clause 2 3 data-privatization-clause 4 data-privatization-in-clause 5 nowait unique-single-clause: 6 copyprivate ( variable-list ) 7 task-construct: 8 task-directive structured-block 9 task-directive: 10 # pragma omp task task-clauseoptseq new-line 11 task-clause: 12 13 unique-task-clause 14 data-default-clause 15 data-privatization-clause 16 data-privatization-in-clause 17 data-sharing-clause unique-task-clause: 18 19 if ( scalar-expression ) 20 final( scalar-expression ) 21 untied 22 mergeable parallel-for-construct: 23 parallel-for-directive iteration-statement 24 parallel-for-directive: 25 # pragma omp parallel for parallel-for-clauseoptseq new-line 26 320 OpenMP API • Version 3.1 July 2011 1 2 parallel-for-clause: unique-parallel-clause 3 unique-for-clause 4 data-default-clause 5 data-privatization-clause 6 data-privatization-in-clause 7 data-privatization-out-clause 8 data-sharing-clause 9 data-reduction-clause 10 parallel-sections-construct: 11 12 13 14 parallel-sections-directive section-scope parallel-sections-directive: # pragma omp parallel sections parallel-sections-clauseoptseq new-line parallel-sections-clause: 15 unique-parallel-clause 16 data-default-clause 17 data-privatization-clause 18 data-privatization-in-clause 19 data-privatization-out-clause 20 data-sharing-clause 21 data-reduction-clause 22 23 24 25 26 27 28 master-construct: master-directive structured-block master-directive: # pragma omp master new-line critical-construct: critical-directive structured-block Appendix C OpenMP C and C++ Grammar 321 critical-directive: 1 # pragma omp critical region-phraseopt new-line 2 region-phrase: 3 ( identifier ) 4 5 barrier-directive: 6 # pragma omp barrier new-line 7 taskwait-directive: 8 # pragma omp taskwait new-line 9 taskyield-directive: 10 # pragma omp taskyield new-line 11 atomic-construct: 12 atomic-directive expression-statement 13 atomic-directive structured block 14 atomic-directive: 15 # pragma omp atomic atomic-clauseopt new-line 16 atomic-clause: 17 18 read 19 write 20 update 21 capture 22 flush-directive: # pragma omp flush flush-varsopt new-line 23 flush-vars: 24 25 ( variable-list ) 26 ordered-construct: ordered-directive structured-block 27 28 322 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 ordered-directive: # pragma omp ordered new-line declaration: /* standard declarations */ threadprivate-directive threadprivate-directive: # pragma omp threadprivate ( variable-list ) new-line data-default-clause: default ( shared ) default ( none ) data-privatization-clause: private ( variable-list ) data-privatization-in-clause: firstprivate ( variable-list ) data-privatization-out-clause: lastprivate ( variable-list ) data-sharing-clause: shared ( variable-list ) data-reduction-clause: reduction ( reduction-operator : variable-list ) reduction-operator: One of: + * - & ^ | && || max min 23 /* in C */ 24 variable-list: 25 identifier 26 variable-list , identifier Appendix C OpenMP C and C++ Grammar 323 1 /* in C++ */ 2 variable-list: id-expression 3 variable-list , id-expression 4 324 OpenMP API • Version 3.1 July 2011 1 APPENDIX D 2 Interface Declarations 3 4 5 6 This appendix gives examples of the C/C++ header file, the Fortran include file and Fortran module that shall be provided by implementations as specified in Chapter 3. It also includes an example of a Fortran 90 generic interface for a library routine. This is a non-normative section, implementation files may differ. 325 1 D.1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 Example of the omp.h Header File #ifndef _OMP_H_DEF #define _OMP_H_DEF /* * define the lock data types */ typedef void *omp_lock_t; typedef void *omp_nest_lock_t; /* * define the schedule kinds */ typedef enum omp_sched_t { omp_sched_static = 1, omp_sched_dynamic = 2, omp_sched_guided = 3, omp_sched_auto = 4 /* , Add vendor specific schedule constants here */ } omp_sched_t; /* * exported OpenMP functions */ #ifdef __cplusplus extern "C" { #endif extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern 326 void int int int int int void int void int int void int int int int int int omp_set_num_threads(int num_threads); omp_get_num_threads(void); omp_get_max_threads(void); omp_get_thread_num(void); omp_get_num_procs(void); omp_in_parallel(void); omp_set_dynamic(int dynamic_threads); omp_get_dynamic(void); omp_set_nested(int nested); omp_get_nested(void); omp_get_thread_limit(void); omp_set_max_active_levels(int max_active_levels); omp_get_max_active_levels(void); omp_get_level(void); omp_get_ancestor_thread_num(int level); omp_get_team_size(int level); omp_get_active_level(void); omp_in_final(void); OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 extern void extern void omp_set_schedule(omp_sched_t kind, int modifier); omp_get_schedule(omp_sched_t *kind, int *modifier); extern extern extern extern extern void void void void int omp_init_lock(omp_lock_t *lock); omp_destroy_lock(omp_lock_t *lock); omp_set_lock(omp_lock_t *lock); omp_unset_lock(omp_lock_t *lock); omp_test_lock(omp_lock_t *lock); extern extern extern extern extern void void void void int omp_init_nest_lock(omp_nest_lock_t *lock); omp_destroy_nest_lock(omp_nest_lock_t *lock); omp_set_nest_lock(omp_nest_lock_t *lock); omp_unset_nest_lock(omp_nest_lock_t *lock); omp_test_nest_lock(omp_nest_lock_t *lock); extern double omp_get_wtime(void); extern double omp_get_wtick(void); #ifdef __cplusplus } #endif #endif Appendix D Interface Declarations 327 1 D.2 2 Example of an Interface Declaration include File omp_lib_kinds.h: 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 integer omp_lock_kind integer omp_nest_lock_kind ! this selects an integer that is large enough to hold a 64 bit integer parameter ( omp_lock_kind = selected_int_kind( 10 ) ) parameter ( omp_nest_lock_kind = selected_int_kind( 10 ) ) integer omp_sched_kind ! this selects an integer that is large enough to hold a 32 bit integer parameter ( omp_sched_kind = selected_int_kind( 8 ) ) integer ( omp_sched_kind ) omp_sched_static parameter ( omp_sched_static = 1 ) integer ( omp_sched_kind ) omp_sched_dynamic parameter ( omp_sched_dynamic = 2 ) integer ( omp_sched_kind ) omp_sched_guided parameter ( omp_sched_guided = 3 ) integer ( omp_sched_kind ) omp_sched_auto parameter ( omp_sched_auto = 4 ) omp_lib.h: 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 ! default integer type assumed below ! default logical type assumed below ! OpenMP API v3.1 include 'omp_lib_kinds.h' integer openmp_version parameter ( openmp_version = 201107 ) external external integer external integer external integer external integer external logical external external logical external external logical external external external 328 omp_set_num_threads omp_get_num_threads omp_get_num_threads omp_get_max_threads omp_get_max_threads omp_get_thread_num omp_get_thread_num omp_get_num_procs omp_get_num_procs omp_in_parallel omp_in_parallel omp_set_dynamic omp_get_dynamic omp_get_dynamic omp_set_nested omp_get_nested omp_get_nested omp_set_schedule omp_get_schedule omp_get_thread_limit OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 integer omp_get_thread_limit external omp_set_max_active_levels external omp_get_max_active_levels integer omp_get_max_active_levels external omp_get_level integer omp_get_level external omp_get_ancestor_thread_num integer omp_get_ancestor_thread_num external omp_get_team_size integer omp_get_team_size external omp_get_active_level integer omp_get_active_level external omp_in_final logical omp_in_final external external external external external logical omp_init_lock omp_destroy_lock omp_set_lock omp_unset_lock omp_test_lock omp_test_lock external external external external external integer omp_init_nest_lock omp_destroy_nest_lock omp_set_nest_lock omp_unset_nest_lock omp_test_nest_lock omp_test_nest_lock external omp_get_wtick double precision omp_get_wtick external omp_get_wtime double precision omp_get_wtime 35 Appendix D Interface Declarations 329 1 D.3 Example of a Fortran Interface Declaration module 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 ! the "!" of this comment starts in column 1 !23456 & & & & module omp_lib_kinds integer, parameter :: omp_lock_kind = selected_int_kind( 10 ) integer, parameter :: omp_nest_lock_kind = selected_int_kind( 10 ) integer, parameter :: omp_sched_kind = selected_int_kind( 8 ) integer(kind=omp_sched_kind), parameter :: omp_sched_static = 1 integer(kind=omp_sched_kind), parameter :: omp_sched_dynamic = 2 integer(kind=omp_sched_kind), parameter :: omp_sched_guided = 3 integer(kind=omp_sched_kind), parameter :: omp_sched_auto = 4 end module omp_lib_kinds module omp_lib use omp_lib_kinds ! OpenMP API v3.1 integer, parameter :: openmp_version = 201107 interface subroutine omp_set_num_threads (number_of_threads_expr) integer, intent(in) :: number_of_threads_expr end subroutine omp_set_num_threads function omp_get_num_threads () integer :: omp_get_num_threads end function omp_get_num_threads function omp_get_max_threads () integer :: omp_get_max_threads end function omp_get_max_threads function omp_get_thread_num () integer :: omp_get_thread_num end function omp_get_thread_num function omp_get_num_procs () integer :: omp_get_num_procs end function omp_get_num_procs function omp_in_parallel () 330 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 logical :: omp_in_parallel end function omp_in_parallel subroutine omp_set_dynamic (enable_expr) logical, intent(in) ::enable_expr end subroutine omp_set_dynamic function omp_get_dynamic () logical :: omp_get_dynamic end function omp_get_dynamic subroutine omp_set_nested (enable_expr) logical, intent(in) :: enable_expr end subroutine omp_set_nested function omp_get_nested () logical :: omp_get_nested end function omp_get_nested subroutine omp_set_schedule (kind, modifier) use omp_lib_kinds integer(kind=omp_sched_kind), intent(in) :: kind integer, intent(in) :: modifier end subroutine omp_set_schedule subroutine omp_get_schedule (kind, modifier) use omp_lib_kinds integer(kind=omp_sched_kind), intent(out) :: kind integer, intent(out)::modifier end subroutine omp_get_schedule function omp_get_thread_limit() integer :: omp_get_thread_limit end function omp_get_thread_limit subroutine omp_set_max_active_levels(var) integer, intent(in) :: var end subroutine omp_set_max_active_levels function omp_get_max_active_levels() integer :: omp_get_max_active_levels end function omp_get_max_active_levels function omp_get_level() integer :: omp_get_level end function omp_get_level function omp_get_ancestor_thread_num(level) integer, intent(in) :: level integer :: omp_get_ancestor_thread_num end function omp_get_ancestor_thread_num Appendix D Interface Declarations 331 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 function omp_get_team_size(level) integer, intent(in) :: level integer :: omp_get_team_size end function omp_get_team_size function omp_get_active_level() integer :: omp_get_active_level end function omp_get_active_level function omp_in_final() logical omp_in_final end function omp_in_final subroutine omp_init_lock (var) use omp_lib_kinds integer (kind=omp_lock_kind), intent(out) :: var end subroutine omp_init_lock subroutine omp_destroy_lock (var) use omp_lib_kinds integer (kind=omp_lock_kind), intent(inout) :: var end subroutine omp_destroy_lock subroutine omp_set_lock (var) use omp_lib_kinds integer (kind=omp_lock_kind), intent(inout) :: var end subroutine omp_set_lock subroutine omp_unset_lock (var) use omp_lib_kinds integer (kind=omp_lock_kind), intent(inout) :: var end subroutine omp_unset_lock function omp_test_lock (var) use omp_lib_kinds logical :: omp_test_lock integer (kind=omp_lock_kind), intent(inout) :: var end function omp_test_lock subroutine omp_init_nest_lock (var) use omp_lib_kinds integer (kind=omp_nest_lock_kind), intent(out) :: var end subroutine omp_init_nest_lock subroutine omp_destroy_nest_lock (var) use omp_lib_kinds integer (kind=omp_nest_lock_kind), intent(inout) :: var end subroutine omp_destroy_nest_lock subroutine omp_set_nest_lock (var) use omp_lib_kinds integer (kind=omp_nest_lock_kind), intent(inout) :: var 332 OpenMP API • Version 3.1 July 2011 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 end subroutine omp_set_nest_lock subroutine omp_unset_nest_lock (var) use omp_lib_kinds integer (kind=omp_nest_lock_kind), intent(inout) :: var end subroutine omp_unset_nest_lock function omp_test_nest_lock (var) use omp_lib_kinds integer :: omp_test_nest_lock integer (kind=omp_nest_lock_kind), intent(inout) :: var end function omp_test_nest_lock function omp_get_wtick () double precision :: omp_get_wtick end function omp_get_wtick function omp_get_wtime () double precision :: omp_get_wtime end function omp_get_wtime end interface end module omp_lib Appendix D Interface Declarations 333 D.4 2 Example of a Generic Interface for a Library Routine 3 4 Any of the OpenMP runtime library routines that take an argument may be extended with a generic interface so arguments of different KIND type can be accommodated. 5 6 The OMP_SET_NUM_THREADS interface could be specified in the omp_lib module as the following: 1 ! the "!" of this comment starts in column 1 interface omp_set_num_threads subroutine omp_set_num_threads_1 ( number_of_threads_expr ) use omp_lib_kinds integer ( kind=selected_int_kind( 8 ) ), intent(in) :: & & number_of_threads_expr end subroutine omp_set_num_threads_1 subroutine omp_set_num_threads_2 ( number_of_threads_expr ) use omp_lib_kinds integer ( kind=selected_int_kind( 3 ) ), intent(in) :: & & number_of_threads_expr end subroutine omp_set_num_threads_2 end interface omp_set_num_threads 7 8 334 OpenMP API • Version 3.1 July 2011 1 APPENDIX E 3 OpenMP ImplementationDefined Behaviors 4 5 6 7 This appendix summarizes the behaviors that are described as implementation defined in this API. Each behavior is cross-referenced back to its description in the main specification. An implementation is required to define and document its behavior in these cases. 2 8 9 10 11 • Memory model: the minimum size at which a memory update may also read and 12 13 14 • Internal control variables: the initial values of nthreads-var, dyn-var, run-sched-var, 15 16 17 18 • Dynamic adjustment of threads: providing the ability to dynamically adjust the 19 20 21 22 • Loop directive: the integer type or kind used to compute the iteration count of a 23 24 • sections construct: the method of scheduling the structured blocks among threads 25 26 • single construct: the method of choosing a thread to execute the structured block 27 28 • Task scheduling points: where task scheduling points occur in untied task regions is write back adjacent variables that are part of another variable (as array or structure elements) is implementation defined but is no larger than required by the base language (see Section 1.4.1 on page 13). def-sched-var, bind-var, stacksize-var, wait-policy-var, thread-limit-var, and maxactive-levels-var are implementation defined (see Section 2.3.2 on page 29). number of threads is implementation defined . Implementations are allowed to deliver fewer threads (but at least one) than indicated in Algorithm 2-1 even if dynamic adjustment is disabled (see Section 2.4.1 on page 36). collapsed loop is implementation defined. The effect of the schedule(runtime) clause when the run-sched-var ICV is set to auto is implementation defined. See Section 2.5.1 on page 39. in the team is implementation defined (see Section 2.5.2 on page 48). is implementation defined (see Section 2.5.3 on page 50). implementation defined (see Section 2.7.3 on page 65). 335 1 2 3 4 • atomic construct: a compliant implementation may enforce exclusive access 5 6 • omp_set_num_threads routine: if the argument is not a positive integer the 7 8 9 • omp_set_schedule routine: for implementation specific schedule types, the between atomic regions which update different storage locations. The circumstances under which this occurs are implementation defined (see Section 2.8.5 on page 73). behavior is implementation defined (see Section 3.2.1 on page 116). values and associated meanings of the second argument are implementation defined. (see Section 3.2.11 on page 128). 10 11 12 13 14 • omp_set_max_active_levels routine: when called from within any explicit 15 16 17 18 • omp_get_max_active_levels routine: when called from within any explicit 19 20 21 • OMP_SCHEDULE environment variable: if the value of the variable does not 22 23 24 25 • OMP_NUM_THREADS environment variable: if any value of the list specified in the 26 27 • OMP_PROC_BIND environment variable: if the value is neither true nor false the 28 29 • OMP_DYNAMIC environment variable: if the value is neither true nor false the 30 31 • OMP_NESTED environment variable: if the value is neither true nor false the 32 33 34 • OMP_STACKSIZE environment variable: if the value does not conform to the 35 36 • OMP_WAIT_POLICY environment variable: the details of the ACTIVE and 37 38 39 40 • OMP_MAX_ACTIVE_LEVELS environment variable: if the value is not a non- parallel region the binding thread set (and binding region, if required) for the omp_set_max_active_levels region is implementation defined and the behavior is implementation defined. If the argument is not a non-negative integer then the behavior is implementation defined (see Section 3.2.14 on page 132). parallel region the binding thread set (and binding region, if required) for the omp_get_max_active_levels region is implementation defined (see Section 3.2.15 on page 134). conform to the specified format then the result is implementation defined (see Section 4.1 on page 154). OMP_NUM_THREADS environment variable leads to a number of threads that is greater than the implementation can support, or if any value is not a positive integer, then the result is implementation defined (see Section 4.2 on page 155). behavior is implementation defined (see Section 4.4 on page 156). behavior is implementation defined (see Section 4.3 on page 156). behavior is implementation defined (see Section 4.5 on page 157). specified format or the implementation cannot provide a stack of the specified size then the behavior is implementation defined (see Section 4.6 on page 157). PASSIVE behaviors are implementation defined (see Section 4.7 on page 158). negative integer or is greater than the number of parallel levels an implementation can support then the behavior is implementation defined (see Section 4.8 on page 159). 336 OpenMP API • Version 3.1 July 2011 1 2 3 4 • OMP_THREAD_LIMIT environment variable: if the requested value is greater than the number of threads an implementation can support, or if the value is not a positive integer, the behavior of the program is implementation defined (see Section 4.9 on page 160). Fortran 5 6 7 8 • threadprivate directive: if the conditions for values of data in the threadprivate 9 10 11 12 13 • shared clause: passing a shared variable to a non-intrinsic procedure may result in 14 15 16 17 18 • Runtime library definitions: it is implementation defined whether the include file objects of threads (other than the initial thread) to persist between two consecutive active parallel regions do not all hold, the allocation status of an allocatable array in the second region is implementation defined (see Section 2.9.2 on page 88). the value of the shared variable being copied into temporary storage before the procedure reference, and back out of the temporary storage into the actual argument storage after the procedure reference. Situations where this occurs other than those specified are implementation defined (see Section 2.9.3.2 on page 94). omp_lib.h or the module omp_lib (or both) is provided. It is implementation defined whether any of the OpenMP runtime library routines that take an argument are extended with a generic interface so arguments of different KIND type can be accommodated (see Section 3.1 on page 114). Fortran Appendix E OpenMP Implementation-Defined Behaviors 337 1 This page intentionally left blank. 2 338 OpenMP API • Version 3.1 July 2011 1 APPENDIX F 2 Features History 3 4 This appendix summarizes the major changes between the OpenMP API Version 2.5 and Version 3.0, and between Version 3.0 and Version 3.1. 5 F.1 Version 3.0 to 3.1 Differences 6 7 • The final and mergeable clauses (see Section 2.7.1 on page 61) were added to 8 9 • The taskyield construct (see Section 2.7.2 on page 64) was added to allow user- the task construct to support optimization of task data environments. defined task switching points. 10 11 12 • The atomic construct (see Section 2.8.5 on page 73) was extended to include 13 14 • Data environment restrictions were changed to allow intent(in) and const- 15 16 17 • Data environment restrictions were changed to allow Fortran pointers in 18 19 • New reduction operators min and max were added for C and C++ (see 20 21 22 23 • The nesting restrictions in Section 2.10 on page 111 were clarified to disallow 24 25 • The omp_in_final runtime library routine (see Section 3.2.20 on page 140) was read, write, and capture forms, and an update clause was added to apply the already existing form of the atomic construct. qualified types for the firstprivate clause (see Section 2.9.3.4 on page 98). firstprivate (see Section 2.9.3.4 on page 98) and lastprivate (see Section 2.9.3.5 on page 101). Section 2.9.3.6 on page 103 and page 105) closely-nested OpenMP regions within an atomic region. This allows an atomic region to be consistently defined with other OpenMP regions so that they include all the code in the atomic construct. added to support specialization of final task regions. 339 1 2 3 4 5 • The nthreads-var ICV has been modified to be a list of the number of threads to use 6 7 8 • The bind-var ICV has been added, which controls whether or not threads are bound 9 • Descriptions of examples (see Appendix A on page 161) were expanded and clarified. at each nested parallel region level. The value of this ICV is still set with the OMP_NUM_THREADS environment variable (see Section 4.2 on page 155), but the algorithm for determining the number of threads used in a parallel region has been modified to handle a list (see Section 2.4.1 on page 36). to processors (see Section 2.3.1 on page 28). The value of this ICV can be set with the OMP_PROC_BIND environment variable (see Section 4.4 on page 156). • Replaced incorrect use of omp_integer_kind in Fortran interfaces (see 10 11 12 13 Section D.3 on page 330 and Section D.4 on page 334) with selected_int_kind(8). F.2 Version 2.5 to 3.0 Differences 14 15 The concept of tasks has been added to the OpenMP execution model (see Section 1.2.3 on page 8 and Section 1.3 on page 12). 16 17 • The task construct (see Section 2.7 on page 61) has been added, which provides a 18 19 • The taskwait construct (see Section 2.8.4 on page 72) has been added, which 20 21 22 • The OpenMP memory model now covers atomicity of memory accesses (see 23 24 25 26 27 28 29 • In Version 2.5, there was a single copy of the nest-var, dyn-var, nthreads-var and 30 31 32 • The definition of active parallel region has been changed: in Version 3.0 a 33 34 • The rules for determining the number of threads used in a parallel region have 35 36 • In Version 3.0, the assignment of iterations to threads in a loop construct with a mechanism for creating tasks explicitly. causes a task to wait for all its child tasks to complete. Section 1.4.1 on page 13). The description of the behavior of volatile in terms of flush was removed. run-sched-var internal control variables (ICVs) for the whole program. In Version 3.0, there is one copy of these ICVs per task (see Section 2.3 on page 28). As a result, the omp_set_num_threads, omp_set_nested and omp_set_dynamic runtime library routines now have specified effects when called from inside a parallel region (see Section 3.2.1 on page 116, Section 3.2.7 on page 123 and Section 3.2.9 on page 125). parallel region is active if it is executed by a team consisting of more than one thread (see Section 1.2.2 on page 2). been modified (see Section 2.4.1 on page 36). static schedule kind is deterministic (see Section 2.5.1 on page 39). 340 OpenMP API • Version 3.1 July 2011 1 2 3 • In Version 3.0, a loop construct may be associated with more than one perfectly 4 5 • Random access iterators, and variables of unsigned integer type, may now be used as 6 7 8 • The schedule kind auto has been added, which gives the implementation the nested loop. The number of associated loops may be controlled by the collapse clause (see Section 2.5.1 on page 39). loop iterators in loops associated with a loop construct (see Section 2.5.1 on page 39). freedom to choose any possible mapping of iterations in a loop construct to threads in the team (see Section 2.5.1 on page 39). 9 10 • Fortran assumed-size arrays now have predetermined data-sharing attributes (see 11 12 • In Fortran, firstprivate is now permitted as an argument to the default 13 14 15 16 17 • For list items in the private clause, implementations are no longer permitted to use 18 19 20 21 22 • In Version 3.0, Fortran allocatable arrays may appear in private, 23 24 • In Version 3.0, static class members variables may appear in a threadprivate 25 26 27 28 • Version 3.0 makes clear where, and with which arguments, constructors and 29 30 31 • The runtime library routines omp_set_schedule and omp_get_schedule 32 33 34 35 36 • The thread-limit-var ICV has been added, which controls the maximum number of 37 38 39 40 • The max-active-levels-var ICV has been added, which controls the number of nested Section 2.9.1.1 on page 84). clause (see Section 2.9.3.1 on page 93). the storage of the original list item to hold the new list item on the master thread. If no attempt is made to reference the original list item inside the parallel region, its value is well defined on exit from the parallel region (see Section 2.9.3.3 on page 96). firstprivate, lastprivate, reduction, copyin and copyprivate clauses. (see Section 2.9.2 on page 88, Section 2.9.3.3 on page 96, Section 2.9.3.4 on page 98, Section 2.9.3.5 on page 101, Section 2.9.3.6 on page 103, Section 2.9.4.1 on page 107 and Section 2.9.4.2 on page 109). directive (see Section 2.9.2 on page 88). destructors of private and threadprivate class type variables are called (see Section 2.9.2 on page 88, Section 2.9.3.3 on page 96, Section 2.9.3.4 on page 98, Section 2.9.4.1 on page 107 and Section 2.9.4.2 on page 109) have been added; these routines respectively set and retrieve the value of the run-sched-var ICV (see Section 3.2.11 on page 128 and Section 3.2.12 on page 130). threads participating in the OpenMP program. The value of this ICV can be set with the OMP_THREAD_LIMIT environment variable and retrieved with the omp_get_thread_limit runtime library routine (see Section 2.3.1 on page 28, Section 3.2.13 on page 131 and Section 4.9 on page 160). active parallel regions. The value of this ICV can be set with the OMP_MAX_ACTIVE_LEVELS environment variable and the omp_set_max_active_levels runtime library routine, and it can be retrieved Appendix F Features History 341 1 2 3 with the omp_get_max_active_levels runtime library routine (see Section 2.3.1 on page 28, Section 3.2.14 on page 132, Section 3.2.15 on page 134 and Section 4.8 on page 159). 4 5 6 7 • The stacksize-var ICV has been added, which controls the stack size for threads that the OpenMP implementation creates. The value of this ICV can be set with the OMP_STACKSIZE environment variable (see Section 2.3.1 on page 28 and Section 4.6 on page 157). 8 9 10 • The wait-policy-var ICV has been added, which controls the desired behavior of 11 12 13 • The omp_get_level runtime library routine has been added, which returns the 14 15 16 • The omp_get_ancestor_thread_num runtime library routine has been added, 17 18 19 • The omp_get_team_size runtime library routine has been added, which returns, 20 21 22 • The omp_get_active_level runtime library routine has been added, which 23 24 • In Version 3.0, locks are owned by tasks, not by threads (see Section 3.3 on page waiting threads. The value of this ICV can be set with the OMP_WAIT_POLICY environment variable (see Section 2.3.1 on page 28 and Section 4.7 on page 158). number of nested parallel regions enclosing the task that contains the call (see Section 3.2.16 on page 135). which returns, for a given nested level of the current thread, the thread number of the ancestor (see Section 3.2.17 on page 136). for a given nested level of the current thread, the size of the thread team to which the ancestor belongs (see Section 3.2.18 on page 137). returns the number of nested, active parallel regions enclosing the task that contains the call (see Section 3.2.19 on page 139). 141). 342 OpenMP API • Version 3.1 July 2011 Index Symbols _OPENMP macro, 2-26 A atomic, 2-73 attributes, data-sharing, 2-84 auto, 2-44 B barrier, 2-70 C capture, atomic, 2-73 clauses collapse, 2-42 copyin, 2-107 copyprivate, 2-109 data-sharing, 2-92 default, 2-93 firstprivate, 2-98 lastprivate, 2-101 private, 2-96 reduction, 2-103 schedule, 2-43 shared, 2-94 collapse, 2-42 compliance, 1-17 conditional compilation, 2-26 constructs atomic, 2-73 barrier, 2-70 critical, 2-68 do, Fortran, 2-41 flush, 2-78 for, C/C++, 2-39 loop, 2-39 master, 2-67 ordered, 2-82 parallel, 2-33 parallel for, C/C++, 2-56 parallel sections, 2-57 parallel workshare, Fortran, 2-59 sections, 2-48 single, 2-50 task, 2-61 taskwait, 2-72 taskyield, 2-64 workshare, 2-52 worksharing, 2-38 copyin, 2-107 copyprivate, 2-109 critical, 2-68 D data sharing, 2-84 data-sharing clauses, 2-92 default, 2-93 directives, 2-21 format, 2-22 threadprivate, 2-88 see also constructs do, Fortran, 2-41 dynamic, 2-44 Index 343 E N environment variables, 4-153 modifying ICV’s, 2-29 OMP_DYNAMIC, 4-156 OMP_MAX_ACTIVE_LEVELS, 4-159 OMP_NESTED, 4-157 OMP_NUM_THREADS, 4-155 OMP_SCHEDULE, 4-154 OMP_STACKSIZE, 4-157 OMP_THREAD_LIMIT, 4-160 OMP_WAIT_POLICY, 4-158 Examples, A-161 execution model, 1-12 nested parallelism, 1-12, 2-28, 3-125 nesting, 2-111 number of threads, 2-36 F firstprivate, 2-98 flush, 2-78 flush operation, 1-15 for, C/C++, 2-39 G glossary, 1-2 grammar rules, C-316 guided, 2-44 H header files, 3-114, D-325 I ICVs (internal control variables), 2-28 implementation, E-335 include files, 3-114, D-325 internal control variables (ICVs), 2-28 L lastprivate, 2-101 loop, scheduling, 2-47 M master, 2-67 memory model, 1-13 model execution, 1-12 memory, 1-13 Index-344 OpenMP API • Version 3.1 July 2011 O omp_destroy_lock, 3-144 omp_destroy_nest_lock, 3-144 OMP_DYNAMIC, 4-156 omp_get_active_level, 3-139 omp_get_ancestor_thread_num, 3-136 omp_get_dynamic, 3-124 omp_get_level, 3-135 omp_get_max_active_levels, 3-134 omp_get_max_threads, 3-118 omp_get_nested, 3-126 omp_get_num_procs, 3-121 omp_get_num_threads, 3-117 omp_get_schedule, 3-130 omp_get_team_size, 3-137 omp_get_thread_limit, 3-131 omp_get_thread_num, 3-119 omp_get_wtick, 3-150 omp_get_wtime, 3-148 omp_in_final, 3-140 omp_in_parallel, 3-122 omp_init_lock, 3-143 omp_init_nest_lock, 3-143 omp_lock_kind, 3-142 omp_lock_t, 3-142 OMP_MAX_ACTIVE_LEVELS, 4-159 omp_nest_lock_kind, 3-142 omp_nest_lock_t, 3-142 OMP_NESTED, 4-157 OMP_NUM_THREADS, 4-155 OMP_SCHEDULE, 4-154 omp_set_dynamic, 3-123 omp_set_lock, 3-145 omp_set_max_active_levels, 3-132 omp_set_nest_lock, 3-145 omp_set_nested, 3-125 omp_set_num_threads, 3-116 omp_set_schedule, 3-128 OMP_STACKSIZE, 4-157 omp_test_lock, 3-147 omp_test_nest_lock, 3-147 OMP_THREAD_LIMIT, 4-160 omp_unset_lock, 3-146 omp_unset_nest_lock, 3-146 OMP_WAIT_POLICY, 4-158 OpenMP compliance, 1-17 examples, A-161 features history, F-339 implementation, E-335 ordered, 2-82 P parallel, 2-33 parallel do, 2-56 parallel for, C/C++, 2-56 parallel sections, 2-57 parallel workshare, Fortran, 2-59 pragmas see constructs private, 2-96 R read, atomic, 2-73 reduction, 2-103 references, 1-17 regions, nesting, 2-111 runtime, 2-45 runtime library interfaces and prototypes, 3-114 synchronization, locks constructs, 2-67 routines, 3-141 T task scheduling, 2-65 task, 2-61 tasking, 2-61 taskwait, 2-72 taskyield, 2-64 terminology, 1-2 threadprivate, 2-88 timer, 3-148 timing routines, 3-148 U update, atomic, 2-73 V variables, environment, 4-153 W wall clock timer, 3-148 website www.openmp.org workshare, 2-52 worksharing constructs, 2-38 parallel, 2-55 scheduling, 2-47 write, atomic, 2-73 S schedule, 2-43 scheduling loop, 2-47 tasks, 2-65 sections, 2-48 shared, 2-94 single, 2-50 static, 2-44 stubs for runtime library routines C/C++, B-302 Fortran, B-309 Index-345 Index-346 OpenMP API • Version 3.1 July 2011