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