OpenMP Application Program Interface Version 4.0 - July 2013 Copyright © 1997-2013 OpenMP Architecture Review Board. Permission to copy without fee all or part of this material is granted, provided the OpenMP Architecture Review Board copyright notice and the title of this document appear. Notice is given that copying is by permission of OpenMP Architecture Review Board. This page is intentionally blank. C O N T E N TS 1. 2. Introduction ...............................................1 1.1 Scope ................................................1 1.2 Glossary ..............................................2 1.2.1 Threading Concepts 1.2.2 OpenMP Language Terminology 1.2.3 Synchronization Terminology 1.2.4 Tasking Terminology 1.2.5 Data Terminology 1.2.6 Implementation Terminology 1.3 Execution Model 1.4 Memory Model ..............................2 .....................2 ........................8 ..............................8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 . . . . . . . . . . . . . . . . . . . . . . . . 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 1.4.1 Structure of the OpenMP Memory Model 1.4.2 Device Data Environments 1.4.3 The Flush Operation 1.4.4 OpenMP Memory Consistency . . . . . . . . . . . . . . . 17 . . . . . . . . . . . . . . . . . . . . . . . . . 18 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 . . . . . . . . . . . . . . . . . . . . . . 20 1.5 OpenMP Compliance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 1.6 Normative References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 1.7 Organization of this document Directives 2.1 2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 Directive Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.1.1 Fixed Source Form Directives . . . . . . . . . . . . . . . . . . . . . . . 27 2.1.2 Free Source Form Directives . . . . . . . . . . . . . . . . . . . . . . . . 28 2.1.3 Stand-Alone Directives Conditional Compilation 2.2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 Fixed Source Form Conditional Compilation Sentinels . . . . 32 i 2.2.2 2.3 . . . . . . 33 Internal Control Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.3.1 ICV Descriptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.3.2 ICV Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.3.3 Modifying and Retrieving ICV Values 2.3.4 How ICVs are Scoped 2.3.5 ICV Override Relationships 2.4 Array Sections 2.5 parallel Construct . . . . . . . . . . . . . . . . . . 37 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 . . . . . . . . . . . . . . . . . . . . . . . . . 40 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 2.5.1 Determining the Number of Threads for a parallel Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 2.5.2 Controlling OpenMP Thread Affinity . . . . . . . . . . . . . . . . . . . 49 2.6 Canonical Loop Form 2.7 Worksharing Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 2.7.1 Loop Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 2.7.2 sections Construct 2.7.3 single Construct 2.7.4 workshare Construct 2.8 2.9 ii Free Source Form Conditional Compilation Sentinel SIMD Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 2.8.1 simd construct 2.8.2 declare simd construct 2.8.3 Loop SIMD construct Device Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 2.9.1 target data Construct 2.9.2 target Construct 2.9.3 target update Construct . . . . . . . . . . . . . . . . . . . . . . . . . 81 2.9.4 declare target Directive . . . . . . . . . . . . . . . . . . . . . . . . . 83 2.9.5 teams Construct 2.9.6 distribute Construct 2.9.7 distribute simd Construct OpenMP API • Version 4.0 - July 2013 . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 . . . . . . . . . . . . . . . . . . . . . . . 91 2.10 2.9.8 Distribute Parallel Loop Construct 2.9.9 Distribute Parallel Loop SIMD Construct Combined Constructs . . . . . . . . . . . . . . . . . . . . 92 . . . . . . . . . . . . . . . 94 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 2.10.1 Parallel Loop Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 2.10.2 parallel sections Construct . . . . . . . . . . . . . . . . . . . . . 97 2.10.3 parallel workshare Construct 2.10.4 Parallel Loop SIMD Construct 2.10.5 target teams construct . . . . . . . . . . . . . . . . . . . . 99 . . . . . . . . . . . . . . . . . . . . . . . 100 . . . . . . . . . . . . . . . . . . . . . . . . . . 101 2.10.6 teams distribute Construct . . . . . . . . . . . . . . . . . . . . . . 102 2.10.7 teams distribute simd Construct . . . . . . . . . . . . . . . . . 104 2.10.8 target teams distribute Construct . . . . . . . . . . . . . . . 105 2.10.9 target teams distribute simd Construct 2.10.10 Teams Distribute Parallel Loop Construct . . . . . . . . . . 106 . . . . . . . . . . . . . . 107 2.10.11 Target Teams Distribute Parallel Loop Construct . . . . . . . . . 109 2.10.12 Teams Distribute Parallel Loop SIMD Construct . . . . . . . . . 110 2.10.13 Target Teams Distribute Parallel Loop SIMD Construct 2.11 Tasking Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 2.11.1 task Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113 2.11.2 taskyield Construct 2.11.3 Task Scheduling 2.12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 Master and Synchronization Constructs 2.12.1 master Construct 2.12.3 barrier Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 2.12.4 taskwait Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 2.12.5 taskgroup Construct 2.12.6 atomic Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 2.12.7 flush Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 2.12.8 ordered Construct Cancellation Constructs . . . . . . . . . . . . . . . . . . . . . . 120 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 2.12.2 critical Construct 2.13 . . . . 111 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 iii 2.13.1 cancel Construct . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 2.13.2 cancellation point Construct 2.14 Data Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 2.14.1 Data-sharing Attribute Rules 2.14.2 threadprivate Directive . . . . . . . . . . . . . . . . . . . . . . . . 146 . . . . . . . . . . . . . . . . . . . . . . . . . 150 2.14.3 Data-Sharing Attribute Clauses 2.14.4 Data Copying Clauses 2.14.5 map Clause 3. . . . . . . . . . . . . . . . . . . . . . . 155 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 2.15 declare reduction Directive 2.16 Nesting of Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 Runtime Library Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 3.1 Runtime Library Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 3.2 Execution Environment Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 3.2.1 omp_set_num_threads . . . . . . . . . . . . . . . . . . . . . . . . . . 189 3.2.2 omp_get_num_threads . . . . . . . . . . . . . . . . . . . . . . . . . . 191 3.2.3 omp_get_max_threads . . . . . . . . . . . . . . . . . . . . . . . . . . 192 3.2.4 omp_get_thread_num 3.2.5 omp_get_num_procs 3.2.6 omp_in_parallel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 3.2.7 omp_set_dynamic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 3.2.8 omp_get_dynamic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 3.2.9 omp_get_cancellation . . . . . . . . . . . . . . . . . . . . . . . . . . . 193 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 . . . . . . . . . . . . . . . . . . . . . . . . . 199 3.2.10 omp_set_nested . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200 3.2.11 omp_get_nested . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 3.2.12 omp_set_schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203 3.2.13 omp_get_schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 3.2.14 omp_get_thread_limit iv . . . . . . . . . . . . . . . . . . . . 143 . . . . . . . . . . . . . . . . . . . . . . . . . 206 3.2.15 omp_set_max_active_levels . . . . . . . . . . . . . . . . . . . . 207 3.2.16 omp_get_max_active_levels . . . . . . . . . . . . . . . . . . . . 209 OpenMP API • Version 4.0 - July 2013 3.2.17 omp_get_level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210 3.2.18 omp_get_ancestor_thread_num 3.2.19 omp_get_team_size . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 3.2.20 omp_get_active_level 3.2.21 omp_in_final . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 3.2.23 omp_set_default_device . . . . . . . . . . . . . . . . . . . . . . . 218 3.2.24 omp_get_default_device . . . . . . . . . . . . . . . . . . . . . . . 219 3.2.25 omp_get_num_devices 3.2.26 omp_get_num_teams 3.2.27 omp_get_team_num . . . . . . . . . . . . . . . . . . . . . . . . . . 220 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222 3.2.28 omp_is_initial_device 3.4 4. . . . . . . . . . . . . . . . . . . . . . . . . . 214 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 3.2.22 omp_get_proc_bind 3.3 . . . . . . . . . . . . . . . . . . 211 Lock Routines . . . . . . . . . . . . . . . . . . . . . . . . 223 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224 3.3.1 omp_init_lock and omp_init_nest_lock 3.3.2 omp_destroy_lock and omp_destroy_nest_lock 3.3.3 omp_set_lock and omp_set_nest_lock 3.3.4 omp_unset_lock and omp_unset_nest_lock 3.3.5 omp_test_lock and omp_test_nest_lock Timing Routines . . . . . . . . . 226 . . . 227 . . . . . . . . . . . . 228 . . . . . . . 229 . . . . . . . . . . 231 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233 3.4.1 omp_get_wtime . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233 3.4.2 omp_get_wtick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234 Environment Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 4.1 OMP_SCHEDULE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238 4.2 OMP_NUM_THREADS 4.3 OMP_DYNAMIC 4.4 OMP_PROC_BIND 4.5 OMP_PLACES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 4.6 OMP_NESTED . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 4.7 OMP_STACKSIZE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 v 4.8 OMP_WAIT_POLICY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245 4.9 OMP_MAX_ACTIVE_LEVELS 4.10 OMP_THREAD_LIMIT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246 4.11 OMP_CANCELLATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246 4.12 OMP_DISPLAY_ENV 4.13 OMP_DEFAULT_DEVICE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248 A. Stubs for Runtime Library Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 A.1 C/C++ Stub Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250 A.2 Fortran Stub Routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257 B. OpenMP C and C++ Grammar B.1 Notation B.2 Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266 C. Interface Declarations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287 C.1 Example of the omp.h Header File . . . . . . . . . . . . . . . . . . . . . . . . . 288 C.2 Example of an Interface Declaration include File . . . . . . . . . . . . . 290 C.3 Example of a Fortran Interface Declaration module . . . . . . . . . . . . 293 C.4 Example of a Generic Interface for a Library Routine . . . . . . . . . . . . 298 D. OpenMP Implementation-Defined Behaviors E. Features History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303 E.1 Version 3.1 to 4.0 Differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303 E.2 Version 3.0 to 3.1 Differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 E.3 Version 2.5 to 3.0 Differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305 Index vi . . . . . . . . . . . . . . . . . . . . . 299 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 OpenMP API • Version 4.0 - July 2013 1 CHAPTER 1 2 Introduction 3 4 5 6 The collection of compiler directives, library routines, and environment variables described in this document collectively define the specification of the OpenMP Application Program Interface (OpenMP API) for shared-memory parallelism in C, C++ and Fortran programs. 7 8 9 10 This specification provides a model for parallel programming that is portable across shared memory architectures from different vendors. Compilers from numerous vendors support the OpenMP API. More information about the OpenMP API can be found at the following web site 11 http://www.openmp.org 12 13 14 15 16 17 18 19 20 The directives, library routines, and environment variables defined in this document allow users to create and manage parallel programs while permitting portability. The directives extend the C, C++ and Fortran base languages with single program multiple data (SPMD) constructs, tasking constructs, device constructs, worksharing constructs, and synchronization constructs, and they provide support for sharing and privatizing data. The functionality to control the runtime environment is provided by library routines and environment variables. Compilers that support the OpenMP API often include a command line option to the compiler that activates and allows interpretation of all OpenMP directives. 21 22 23 24 25 26 27 1.1 Scope The OpenMP API covers only user-directed parallelization, wherein the programmer explicitly specifies the actions to be taken by the compiler and runtime system in order to execute the program in parallel. OpenMP-compliant implementations are not required to check for data dependencies, data conflicts, race conditions, or deadlocks, any of which may occur in conforming programs. In addition, compliant implementations are not required to check for code sequences that cause a program to be classified as non- 1 conforming. Application developers are responsible for correctly using the OpenMP API to produce a conforming program. The OpenMP API does not cover compiler-generated automatic parallelization and directives to the compiler to assist such parallelization. 1 2 3 4 1.2 Glossary 5 1.2.1 Threading Concepts 6 7 8 thread 9 OpenMP thread 10 11 thread-safe routine 12 13 processor 14 device 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). Implementation defined hardware unit on which one or more OpenMP threads can execute. An implementation defined logical execution engine. COMMENT: A device could have one or more processors. 15 16 host device 17 target device 18 An execution entity with a stack and associated static memory, called threadprivate memory. 1.2.2 19 20 21 The device on which the OpenMP program begins execution A device onto which code and data may be offloaded from the host device. OpenMP Language Terminology base language A programming language that serves as the foundation of the OpenMP specification. COMMENT: See Section 1.6 on page 22 for a listing of current base languages for the OpenMP API. 22 23 base program 24 2 A program written in a base language. OpenMP API • Version 4.0 - July 2013 1 2 structured block 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. 3 4 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. 5 COMMENTS: 6 For all base languages, 7 • Access to the structured block must not be the result of a branch. 8 9 • The point of exit cannot be a branch out of the structured block. 10 For C/C++: 11 • The point of entry must not be a call to setjmp(). 12 • longjmp() and throw() must not violate the entry/exit criteria. 13 • Calls to exit() are allowed in a structured block. 14 15 16 17 18 • 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. 19 For Fortran: 20 21 22 • enclosing context In C/C++, the innermost scope enclosing an OpenMP directive. In Fortran, the innermost scoping unit enclosing an OpenMP directive. 23 24 25 STOP statements are allowed in a structured block. directive In C/C++, a #pragma, and in Fortran, a comment, that specifies OpenMP program behavior. COMMENT: See Section 2.1 on page 26 for a description of OpenMP directive syntax. 26 27 28 white space 29 30 OpenMP program 31 32 conforming program 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. An OpenMP program that follows all the rules and restrictions of the OpenMP specification. Chapter 1 Introduction 3 1 2 3 declarative directive An OpenMP directive that may only be placed in a declarative context. A declarative directive results in one or more declarations only; it is not associated with the immediate execution of any user code. 4 5 executable directive An OpenMP directive that is not declarative. That is, it may be placed in an executable context. 6 stand-alone directive 7 8 loop directive 9 associated loop(s) An OpenMP executable directive that has no associated executable user code. An OpenMP executable directive whose associated user code must be a loop nest that is a structured block. 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. 10 11 12 13 14 15 construct 16 17 18 19 20 21 22 23 region An OpenMP executable directive (and for Fortran, the paired end directive, if any) and the associated statement, loop or structured block, if any, not including the code in any called routines. That is, in the lexical extent of an executable directive. 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. The point where a target or teams directive is encountered is a part of the region of the encountering thread, but the region associated with the target or teams directive is not. COMMENTS: 24 25 26 A region may also be thought of as the dynamic or runtime extent of a construct or of an OpenMP library routine. 27 28 During the execution of an OpenMP program, a construct may give rise to many regions. 29 active parallel region 30 31 inactive parallel region 32 4 A parallel region that is executed by a team consisting of more than one thread. A parallel region that is executed by a team of only one thread. OpenMP API • Version 4.0 - July 2013 1 2 3 sequential part All code encountered during the execution of an initial task region that is not part of a parallel region corresponding to a parallel construct or a task region corresponding to a task construct. COMMENTS: 4 5 A sequential part is enclosed by an implicit parallel region. 6 7 8 Executable statements in called routines may be in both a sequential part and any number of explicit parallel regions at different points in the program execution. 9 10 master thread The thread that encounters a parallel construct, creates a team, generates a set of implicit tasks, then executes one of those tasks as thread number 0. 11 12 13 14 15 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. 16 17 18 19 child thread When a thread encounters a parallel construct, each of the threads in the generated parallel region's team are child threads of the encountering thread. The target or teams region's initial thread is not a child thread of the thread that encountered the target or teams construct. 20 ancestor thread For a given thread, its parent thread or one of its parent thread’s ancestor threads. 21 22 descendent thread For a given thread, one of its child threads or one of its child threads’ descendent threads. 23 24 team A set of one or more threads participating in the execution of a parallel region. COMMENTS: 25 26 27 For an active parallel region, the team comprises the master thread and at least one additional thread. 28 29 For an inactive parallel region, the team comprises only the master thread. 30 league 31 contention group 32 33 34 implicit parallel region The set of thread teams created by a target construct or a teams construct. An initial thread and its descendent threads. An inactive parallel region that generates an initial task region. Implicit parallel regions surround the whole OpenMP program, all target regions, and all teams regions. Chapter 1 Introduction 5 1 initial thread 2 nested construct closely nested construct 3 4 nested region 5 6 A thread that executes an implicit parallel region. A construct (lexically) enclosed by another construct. A construct nested inside another construct with no other construct nested between them. A region (dynamically) enclosed by another region. That is, a region encountered during the execution of another region. COMMENT: Some nestings are conforming and some are not. See Section 2.16 on page 186 for the restrictions on nesting. 7 8 9 10 closely nested region 11 all threads 12 current team 13 encountering thread 14 all tasks 15 16 17 current team tasks 18 generating task 19 20 binding thread set A region nested inside another region with no parallel region nested between them. All OpenMP threads participating in the OpenMP program. All threads in the team executing the innermost enclosing parallel region. For a given region, the thread that encounters the corresponding construct. All tasks participating in the OpenMP program. All tasks encountered by the corresponding team. Note that the implicit tasks constituting the parallel region and any descendent tasks encountered during the execution of these implicit tasks are included in this set of tasks. 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. 21 22 The binding thread set for a given region can be all threads on a device, all threads in a contention group, the current team, or the encountering thread. 23 24 COMMENT: The binding thread set for a particular region is described in its corresponding subsection of this specification. binding task set 25 26 The set of tasks that are affected by, or provide the context for, the execution of a region. 27 28 The binding task set for a given region can be all tasks, the current team tasks, or the generating task. 29 30 COMMENT: The binding task set for a particular region (if applicable) is described in its corresponding subsection of this specification. 6 OpenMP API • Version 4.0 - July 2013 1 2 binding region The enclosing region that determines the execution context and limits the scope of the effects of the bound region is called the binding region. 3 4 5 Binding region is not defined for regions whose binding thread set is all threads or the encountering thread, nor is it defined for regions whose binding task set is all tasks. 6 COMMENTS: 7 8 The binding region for an ordered region is the innermost enclosing loop region. 9 10 The binding region for a taskwait region is the innermost enclosing task region. 11 12 13 For all other regions for which the binding thread set is the current team or the binding task set is the current team tasks, the binding region is the innermost enclosing parallel region. 14 15 For regions for which the binding task set is the generating task, the binding region is the region of the generating task. 16 17 A parallel region need not be active nor explicit to be a binding region. 18 A task region need not be explicit to be a binding region. 19 20 A region never binds to any region outside of the innermost enclosing parallel region. 21 22 orphaned construct A construct that gives rise to a region whose binding thread set is the current team, but is not nested within another construct giving rise to the binding region. 23 24 worksharing construct 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. 25 For C/C++, worksharing constructs are for, sections, and single. 26 27 For Fortran, worksharing constructs are do, sections, single and workshare. 28 sequential loop 29 30 place 31 32 place list A loop that is not associated with any OpenMP loop directive. Unordered set of processors that is treated by the execution environment as a location unit when dealing with OpenMP thread affinity. The ordered list that describes all OpenMP places available to the execution environment. Chapter 1 Introduction 7 1 2 3 place partition 4 SIMD instruction 5 6 SIMD lane A software or hardware mechanism capable of processing one data element from a SIMD instruction. 7 8 SIMD chunk A set of iterations executed concurrently, each by a SIMD lane, by a single thread by means of SIMD instructions. 9 SIMD loop 10 1.2.3 A single machine instruction that can can operate on multiple data elements. A loop that includes at least one SIMD chunk. Synchronization Terminology 11 12 13 14 15 barrier 16 17 cancellation 18 19 cancellation point 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. If cancellation has been requested, threads may proceed to the end of the canceled region even if some threads in the team have not reached the barrier. An action that cancels (that is, aborts) an OpenMP region and causes executing implicit or explicit tasks to proceed to the end of the canceled region. A point at which implicit and explicit tasks check if cancellation has been requested. If cancellation has been observed, they perform the cancellation. COMMENT: For a list of cancellation points, see Section 2.13.1 on page 140. 20 21 An ordered list that corresponds to a contiguous interval in the OpenMP place list. It describes the places currently available to the execution environment for a given parallel region. 1.2.4 Tasking Terminology 22 23 task 24 task region A specific instance of executable code and its data environment, generated when a thread encounters a task construct or a parallel construct. 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. 25 26 explicit task A task generated when a task construct is encountered during execution. 27 28 implicit task A task generated by an implicit parallel region or generated when a parallel construct is encountered during execution. 29 initial task 30 31 current task 8 An implicit task associated with an implicit parallel region. For a given thread, the task corresponding to the task region in which it is executing. OpenMP API • Version 4.0 - July 2013 1 2 child task 3 sibling tasks 4 5 descendent task A task that is the child task of a task region or of one of its descendent task regions. 6 7 task completion Task completion occurs when the end of the structured block associated with the construct that generated the task is reached. Tasks that are child tasks of the same task region. COMMENT: Completion of the initial task occurs at program exit. 8 9 10 11 A task is a child task of its generating task region. A child task region is not part of its generating task region. task scheduling point A point during the execution of the current task region at which it can be suspended to be resumed later; or the point of task completion, after which the executing thread may switch to a different task region. COMMENT: For a list of task scheduling points, see Section 2.11.3 on page 118. 12 13 task switching The act of a thread switching from the execution of one task to another task. 14 15 tied task 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. 16 17 untied task A task that, when its task region is suspended, can be resumed by any thread in the team. That is, the task is not tied to any thread. 18 19 20 undeferred task 21 22 23 included task 24 25 merged task 26 final task 27 28 29 task dependence An ordering relation between two sibling tasks: the dependent task and a previously generated predecessor task. The task dependence is fulfilled when the predecessor task has completed. 30 31 dependent task A task that because of a task dependence cannot be executed until its predecessor tasks have completed. 32 predecessor task 33 task synchronization construct A task for which execution is not deferred with respect to its generating task region. That is, its generating task region is suspended until execution of the undeferred task is completed. A task for which execution is sequentially included in the generating task region. That is, an included task is undeferred and executed immediately by the encountering thread. A task whose data environment, inclusive of ICVs, is the same as that of its generating task region. A task that forces all of its child tasks to become final and included tasks. A task that must complete before its dependent tasks can be executed. A taskwait, taskgroup, or a barrier construct. Chapter 1 Introduction 9 1 1.2.5 Data Terminology variable 2 3 Note – An array or structure element is a variable that is part of another variable. 4 5 array section 6 array item 7 8 9 private variable A designated subset of the elements of an array. An array, an array section or an array element. With respect to a given set of task regions or SIMD lanes that bind to the same parallel region, a variable whose name provides access to a different block of storage for each task region or SIMD lane. A variable that is part of another variable (as an array or structure element) cannot be made private independently of other components. 10 11 shared variable 12 13 14 With respect to a given set of task regions that bind to the same parallel region, a variable whose name provides access to the same block of storage for each task region. A variable that is part of another variable (as an array or structure element) cannot be shared independently of the other components, except for static data members of C++ classes. 15 16 17 18 19 20 A named data storage block, whose value can be defined and redefined during the execution of a program. threadprivate variable A variable that is replicated, one instance per thread, by the OpenMP implementation. Its name then provides access to a different block of storage for each thread. A variable that is part of another variable (as an array or structure element) cannot be made threadprivate independently of the other components, except for static data members of C++ classes. 21 22 23 24 threadprivate memory The set of threadprivate variables associated with each thread. 25 data environment The variables associated with the execution of a given region. 26 device data environment A data environment defined by a target data or target construct. 27 10 OpenMP API • Version 4.0 - July 2013 1 2 mapped variable COMMENT: The original and corresponding variables may share storage. 3 4 5 6 7 An original variable in a data environment with a corresponding variable in a device data environment. mappable type A type that is valid for a mapped variable. If a type is composed from other types (such as the type of an array or structure element) and any of the other types are not mappable then the type is not mappable. COMMENT: Pointer types are mappable but the memory block to which the pointer refers is not mapped. 8 9 10 11 For C: The type must be a complete type. 12 13 For C++: The type must be a complete type. 14 In addition, for class types: 15 16 • All member functions accessed in any target region must appear in a declare target directive. 17 • All data members must be non-static. 18 19 20 21 • A mappable type cannot contain virtual members. 22 For Fortran: The type must be definable. defined For variables, the property of having a valid value. 23 24 For C: For the contents of variables, the property of having a valid value. 25 26 27 For C++: For the contents of variables of POD (plain old data) type, the property of having a valid value. 28 29 For variables of non-POD class type, the property of having been constructed but not subsequently destructed. 30 31 32 For Fortran: 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. 33 34 COMMENT: Programs that rely upon variables that are not defined are non-conforming programs. 35 class type For C++: Variables declared with one of the class, struct, or union keywords. Chapter 1 Introduction 11 1 sequentially consistent atomic construct 2 non-sequentially consistent atomic construct 3 1.2.6 6 supporting the OpenMP API 7 supporting nested parallelism internal control variable 8 9 An atomic construct for which the seq_cst clause is not specified. Implementation Terminology supporting n levels of parallelism 4 5 An atomic construct for which the seq_cst clause is specified. 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 runtime 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. 10 11 compliant implementation 12 13 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. 14 15 unspecified behavior 16 17 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. 18 Such unspecified behavior may result from: 19 20 • Issues documented by the OpenMP specification as having unspecified behavior. 21 • A non-conforming program. 22 • A conforming program exhibiting an implementation defined behavior. 12 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 implementation defined Behavior that must be documented by the implementation, and is allowed to vary among different compliant implementations. An implementation is allowed to define this behavior as unspecified. COMMENT: All features that have implementation defined behavior are documented in Appendix D. Chapter 1 Introduction 13 1 1.3 Execution Model 2 3 4 5 6 7 8 9 10 11 12 13 The OpenMP API uses the fork-join model of parallel execution. Multiple threads of execution perform tasks defined implicitly or explicitly by OpenMP directives. The OpenMP API is intended to support programs that will execute correctly both as parallel programs (multiple threads of execution and a full OpenMP support library) and as sequential programs (directives ignored and a simple OpenMP stubs library). However, it is possible and permitted to develop a program that executes correctly as a parallel program but not as a sequential program, or that produces different results when executed as a parallel program compared to when it is executed as a sequential program. Furthermore, using different numbers of threads may result in different numeric results because of changes in the association of numeric operations. For example, a serial addition reduction may have a different pattern of addition associations than a parallel reduction. These different associations may change the results of floating-point addition. 14 15 16 17 An OpenMP program begins as a single thread of execution, called an initial thread. An initial thread executes sequentially, as if enclosed in an implicit task region, called an initial task region, that is defined by the implicit parallel region surrounding the whole program. 18 19 20 21 22 23 24 The thread that executes the implicit parallel region that surrounds the whole program executes on the host device. An implementation may support other target devices. If supported, one or more devices are available to the host device for offloading code and data. Each device has its own threads that are distinct from threads that execute on another device. Threads cannot migrate from one device to another device. The execution model is host-centric such that the host device offloads target regions to target devices. 25 26 27 28 The initial thread that executes the implicit parallel region that surrounds the target region may execute on a target devce. An initial thread executes sequentially, as if enclosed in an implicit task region, called an initial task region, that is defined by an implicit inactive parallel region that surrounds the entire target region. 29 30 31 32 33 When a target construct is encountered, the target region is executed by the implicit device task. The task that encounters the target construct waits at the end of the construct until execution of the region completes. If a target device does not exist, or the target device is not supported by the implementation, or the target device cannot execute the target construct then the target region is executed by the host device. 34 35 36 37 The teams construct creates a league of thread teams where the master thread of each team executes the region. Each of these master threads is an initial thread, and executes sequentially, as if enclosed in an implicit task region that is defined by an implicit parallel region that surrounds the entire teams region. 14 OpenMP API • Version 4.0 - July 2013 1 2 3 If a construct creates a data environment, the data environment is created at the time the construct is encountered. Whether a construct creates a data environment is defined in the description of the construct. 4 5 6 7 8 9 10 11 12 13 14 When any thread encounters a parallel construct, the thread creates a team of itself and zero or more additional threads and becomes the master of the new team. A set of implicit tasks, one per thread, is generated. The code for each task is defined by the code inside the parallel construct. Each task is assigned to a different thread in the team and becomes tied; that is, it is always executed by the thread to which it is initially assigned. The task region of the task being executed by the encountering thread is suspended, and each member of the new team executes its implicit task. There is an implicit barrier at the end of the parallel construct. Only the master thread resumes execution beyond the end of the parallel construct, resuming the task region that was suspended upon encountering the parallel construct. Any number of parallel constructs can be specified in a single program. 15 16 17 18 19 20 21 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. A parallel construct may include a proc_bind clause to specify the places to use for the threads in the team within the parallel region. 22 23 24 25 26 When any team encounters a worksharing construct, the work inside the construct is divided among the members of the team, and executed cooperatively instead of being executed by every thread. There is a default barrier at the end of each worksharing construct unless the nowait clause is present. Redundant execution of code by every thread in the team resumes after the end of the worksharing construct. 27 28 29 30 31 32 33 34 35 36 37 38 39 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 according to task scheduling constraints and thread availability. Threads are allowed to suspend the current task region at a task scheduling point in order to execute a different task. If the suspended task region is for a tied task, the initially assigned thread later resumes execution of the suspended task region. If the suspended task region is for an untied task, then any thread may resume its execution. Completion of all explicit tasks bound to a given parallel region is guaranteed before the master thread leaves the implicit barrier at the end of the region. Completion of a subset of all explicit tasks bound to a given parallel region may be specified through the use of task synchronization constructs. Completion of all explicit tasks bound to the implicit parallel region is guaranteed by the time the program exits. 40 41 42 When any thread encounters a simd construct, the iterations of the loop associated with the construct may be executed concurrently using the SIMD lanes that are available to the thread. Chapter 1 Introduction 15 1 2 3 4 5 6 7 8 9 The cancel construct can alter the previously described flow of execution in an OpenMP region. The effect of the cancel construct depends on its construct-typeclause. If a task encounters a cancel construct with a taskgroup construct-typeclause, then the task activates cancellation and continues execution at the end of its task region, which implies completion of that task. Any other task in that taskgroup that has begun executing completes execution unless it encounters a cancellation point construct, in which case it continues execution at the end of its task region, which implies its completion. Other tasks in that taskgroup region that have not begun execution are aborted, which implies their completion. 10 11 12 13 14 For all other construct-type-clause values, if a thread encounters a cancel construct, it activates cancellation of the innermost enclosing region of the type specified and the thread continues execution at the end of that region. Threads check if cancellation has been activated for their region at cancellation points and, if so, also resume execution at the end of the canceled region. 15 16 17 18 If cancellation has been activated regardless of construct-type-clause, threads that are waiting inside a barrier other than an implicit barrier at the end of the canceled region exit the barrier and resume execution at the end of the canceled region. This action can occur before the other threads reach that barrier. 19 20 21 22 Synchronization constructs and library routines are available in the OpenMP API to coordinate tasks and data access in parallel regions. In addition, library routines and environment variables are available to control or to query the runtime environment of OpenMP programs. 23 24 25 26 27 The OpenMP specification makes no guarantee that input or output to the same file is synchronous when executed in parallel. In this case, the programmer is responsible for synchronizing input and output statements (or routines) using the provided synchronization constructs or library routines. For the case where each thread accesses a different file, no synchronization by the programmer is necessary. 28 16 OpenMP API • Version 4.0 - July 2013 1 1.4 Memory Model 2 1.4.1 Structure of the OpenMP Memory Model 3 4 5 6 7 8 9 10 11 12 The OpenMP API provides a relaxed-consistency, shared-memory model. All OpenMP threads have access to a place to store and to retrieve variables, called the memory. In addition, each thread is allowed to have its own temporary view of the memory. The temporary view of memory for each thread is not a required part of the OpenMP memory model, but can represent any kind of intervening structure, such as machine registers, cache, or other local storage, between the thread and the memory. The temporary view of memory allows the thread to cache variables and thereby to avoid going to memory for every reference to a variable. Each thread also has access to another type of memory that must not be accessed by other threads, called threadprivate memory. 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 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 or SIMD lane 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 2.14.3.3 on page 159 for additional details. References to a private variable in the structured block refer to the private version of the original variable for the current task or SIMD lane. The relationship between the value of the original variable and the initial or final value of the private version depends on the exact clause that specifies it. Details of this issue, as well as other issues with privatization, are provided in Section 2.14 on page 146. 29 30 31 The minimum size at which a memory update may also read and write back adjacent variables that are part of another variable (as array or structure elements) is implementation defined but is no larger than required by the base language. 32 33 34 35 36 37 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. Chapter 1 Introduction 17 1 2 3 4 5 6 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. 7 8 9 10 11 12 13 A private variable in a task region that eventually generates an inner nested parallel region is permitted to be made shared by implicit tasks in the inner parallel region. A private variable in a task region can be shared by an explicit task region generated during its execution. However, it is the programmer’s responsibility to ensure through synchronization that the lifetime of the variable does not end before completion of the explicit task region sharing it. Any other access by one task to the private variables of another task results in unspecified behavior. 14 1.4.2 Device Data Environments 15 16 17 18 19 20 When an OpenMP program begins, each device has an initial device data environment. The initial device data environment for the host device is the data environment associated with the initial task region. Directives that accept data-mapping attribute clauses determine how an original variable is mapped to a corresponding variable in a device data environment. The original variable is the variable with the same name that exists in the data environment of the task that encounters the directive. 21 22 23 24 25 26 27 If a corresponding variable is present in the enclosing device data environment, the new device data environment inherits the corresponding variable from the enclosing device data environment. If a corresponding variable is not present in the enclosing device data environment, a new corresponding variable (of the same type and size) is created in the new device data environment. In the latter case, the initial value of the new corresponding variable is determined from the clauses and the data environment of the encountering thread. 28 29 30 31 32 The corresponding variable in the device data environment may share storage with the original variable. Writes to the corresponding variable may alter the value of the original variable. The impact of this on memory consistency is discussed in Section 1.4.4 on page 20. When a task executes in the context of a device data environment, references to the original variable refer to the corresponding variable in the device data environment. 33 34 35 The relationship between the value of the original variable and the initial or final value of the corresponding variable depends on the map-type. Details of this issue, as well as other issues with mapping a variable, are provided in Section 2.14.5 on page 177. 36 37 38 The original variable in a data environment and the corresponding variable(s) in one or more device data environments may share storage. Without intervening synchronization data races can occur. 18 OpenMP API • Version 4.0 - July 2013 1 1.4.3 The Flush Operation 2 3 4 5 6 7 The memory model has relaxed-consistency because a thread’s temporary view of memory is not required to be consistent with memory at all times. A value written to a variable can remain in the thread’s temporary view until it is forced to memory at a later time. Likewise, a read from a variable may retrieve the value from the thread’s temporary view, unless it is forced to read from memory. The OpenMP flush operation enforces consistency between the temporary view and memory. 8 9 10 11 12 The flush operation is applied to a set of variables called the flush-set. The flush operation restricts reordering of memory operations that an implementation might otherwise do. Implementations must not reorder the code for a memory operation for a given variable, or the code for a flush operation for the variable, with respect to a flush operation that refers to the same variable. 13 14 15 16 17 18 19 20 21 22 23 24 25 26 If a thread has performed a write to its temporary view of a shared variable since its last flush of that variable, then when it executes another flush of the variable, the flush does not complete until the value of the variable has been written to the variable in memory. If a thread performs multiple writes to the same variable between two flushes of that variable, the flush ensures that the value of the last write is written to the variable in memory. A flush of a variable executed by a thread also causes its temporary view of the variable to be discarded, so that if its next memory operation for that variable is a read, then the thread will read from memory when it may again capture the value in the temporary view. When a thread executes a flush, no later memory operation by that thread for a variable involved in that flush is allowed to start until the flush completes. The completion of a flush of a set of variables executed by a thread is defined as the point at which all writes to those variables performed by the thread before the flush are visible in memory to all other threads and that thread’s temporary view of all variables involved is discarded. 27 28 29 30 31 32 The flush operation provides a guarantee of consistency between a thread’s temporary view and memory. Therefore, the flush operation can be used to guarantee that a value written to a variable by one thread may be read by a second thread. To accomplish this, the programmer must ensure that the second thread has not written to the variable since its last flush of the variable, and that the following sequence of events happens in the specified order: 33 1. The value is written to the variable by the first thread. 34 2. The variable is flushed by the first thread. 35 3. The variable is flushed by the second thread. 36 4. The value is read from the variable by the second thread. Chapter 1 Introduction 19 Note – OpenMP synchronization operations, described in Section 2.12 on page 120 and in Section 3.3 on page 224, are recommended for enforcing this order. Synchronization through variables is possible but is not recommended because the proper timing of flushes is difficult. 1 2 3 4 5 1.4.4 OpenMP Memory Consistency The restrictions in Section 1.4.3 on page 19 on reordering with respect to flush operations guarantee the following: 6 7 8 9 10 • If the intersection of the flush-sets of two flushes performed by two different threads 11 12 13 • If two operations performed by the same thread either access, modify, or flush the 14 15 • If the intersection of the flush-sets of two flushes is empty, the threads can observe 16 17 The flush operation can be specified using the flush directive, and is also implied at various locations in an OpenMP program: see Section 2.12.7 on page 134 for details. 18 19 20 Note – Since flush operations by themselves cannot prevent data races, explicit flush operations are only useful in combination with non-sequentially consistent atomic directives. 21 OpenMP programs that: 22 • do not use non-sequentially consistent atomic directives, 23 24 • do not rely on the accuracy of a false result from omp_test_lock and 25 • correctly avoid data races as required in Section 1.4.1 on page 17 26 27 28 29 behave as though operations on shared variables were simply interleaved in an order consistent with the order in which they are performed by each thread. The relaxed consistency model is invisible for such programs, and any explicit flush operations in such programs are redundant. is non-empty, then the two flushes must be completed as if in some sequential order, seen by all threads. same variable, then they must be completed as if in that thread's program order, as seen by all threads. these flushes in any order. omp_test_nest_lock, and 20 OpenMP API • Version 4.0 - July 2013 Implementations are allowed to relax the ordering imposed by implicit flush operations when the result is only visible to programs using non-sequentially consistent atomic directives. 1 2 3 4 1.5 OpenMP Compliance 5 6 7 8 9 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 23) are for information purposes only and are not part of the specification. 10 11 12 13 14 15 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. 16 17 18 19 20 21 All library, intrinsic and built-in routines provided by the base language must be threadsafe in a compliant implementation. In addition, the implementation of the base language must also be thread-safe. For example, ALLOCATE and DEALLOCATE statements must be thread-safe in Fortran. Unsynchronized concurrent use of such routines by different threads must produce correct results (although not necessarily the same as serial execution results, as in the case of random number generation routines). 22 23 24 25 Starting with Fortran 90, 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. 26 27 28 Appendix D 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 D. Chapter 1 Introduction 21 1 1.6 Normative References 2 • ISO/IEC 9899:1990, Information Technology - Programming Languages - C. 3 This OpenMP API specification refers to ISO/IEC 9899:1990 as C90. 4 5 • ISO/IEC 9899:1999, Information Technology - Programming Languages - C. 6 This OpenMP API specification refers to ISO/IEC 9899:1999 as C99. 7 8 • ISO/IEC 14882:1998, Information Technology - Programming Languages - C++. 9 This OpenMP API specification refers to ISO/IEC 14882:1998 as C++. 10 11 • ISO/IEC 1539:1980, Information Technology - Programming Languages - Fortran. 12 This OpenMP API specification refers to ISO/IEC 1539:1980 as Fortran 77. 13 14 • ISO/IEC 1539:1991, Information Technology - Programming Languages - Fortran. 15 This OpenMP API specification refers to ISO/IEC 1539:1991 as Fortran 90. 16 17 • ISO/IEC 1539-1:1997, Information Technology - Programming Languages - Fortran. 18 This OpenMP API specification refers to ISO/IEC 1539-1:1997 as Fortran 95. 19 20 21 • ISO/IEC 1539-1:2004, Information Technology - Programming Languages - Fortran. 22 23 This OpenMP API specification refers to ISO/IEC 1539-1:2004 as Fortran 2003. The following features are not supported: 24 • IEEE Arithmetic issues covered in Fortran 2003 Section 14 25 • Allocatable enhancement 26 • Parameterized derived types 27 • Finalization 28 • Procedures bound by name to a type 29 • The PASS attribute 22 OpenMP API • Version 4.0 - July 2013 1 • Procedures bound to a type as operators 2 • Type extension 3 • Overriding a type-bound procedure 4 • Polymorphic entities 5 • SELECT TYPE construct 6 • Deferred bindings and abstract types 7 • Controlling IEEE underflow 8 • Another IEEE class value Where this OpenMP API specification refers to C, C++ or Fortran, reference is made to the base language supported by the implementation. 9 10 11 1.7 Organization of this document 12 The remainder of this document is structured as follows: 13 • Chapter 2 “Directives” 14 • Chapter 3 “Runtime Library Routines” 15 • Chapter 4 “Environment Variables” 16 • Appendix A “Stubs for Runtime Library Routines” 17 • Appendix B “OpenMP C and C++ Grammar” 18 • Appendix C “Interface Declarations” 19 • Appendix D “OpenMP Implementation-Defined Behaviors” 20 • Appendix E “Features History” 21 22 23 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: 24 C/C++ specific text... C/C++ C/C++ 25 Text that applies only to programs whose base language is C only is shown as follows: 26 C specific text... C C Chapter 1 Introduction 23 Text that applies only to programs whose base language is C90 only is shown as follows: 1 2 C90 C90 specific text... 3 C90 Text that applies only to programs whose base language is C99 only is shown as follows: 4 5 C99 C99 specific text... 6 C99 Text that applies only to programs whose base language is C++ only is shown as follows: 7 8 C++ C++ specific text... 9 C++ Text that applies only to programs whose base language is Fortran is shown as follows: 10 Fortran Fortran specific text...... 11 Fortran Where an entire page consists of, for example, Fortran specific text, a marker is shown at the top of the page like this: 12 13 Fortran (cont.) 14 15 Some text is for information only, and is not part of the normative specification. Such text is designated as a note, like this: 16 Note – Non-normative text.... 24 OpenMP API • Version 4.0 - July 2013 1 CHAPTER 2 2 Directives 3 4 This chapter describes the syntax and behavior of OpenMP directives, and is divided into the following sections: 5 • The language-specific directive format (Section 2.1 on page 26) 6 • Mechanisms to control conditional compilation (Section 2.2 on page 32) 7 8 • How to specify and to use array sections for all base languages (Section 2.4 on page 9 • Control of OpenMP API ICVs (Section 2.3 on page 34) 42) 10 11 • Details of each OpenMP directive (Section 2.5 on page 44 to Section 2.16 on page 12 13 In C/C++, OpenMP directives are specified by using the #pragma mechanism provided by the C and C++ standards. 186) C/C++ C/C++ Fortran 14 15 16 In Fortran, OpenMP directives are specified by using special comments that are identified by unique sentinels. Also, a special comment form is available for conditional compilation. Fortran 17 18 19 20 21 22 Compilers can therefore ignore OpenMP directives and conditionally compiled code if support of the OpenMP API is not provided or enabled. A compliant implementation must provide an option or interface that ensures that underlying support of all OpenMP directives and OpenMP conditional compilation mechanisms is enabled. In the remainder of this document, the phrase OpenMP compilation is used to mean a compilation with these OpenMP features enabled. 25 Fortran 1 Restrictions 2 The following restriction applies to all OpenMP directives: 3 • OpenMP directives may not appear in PURE or ELEMENTAL procedures. Fortran 4 2.1 Directive Format C/C++ OpenMP directives for C/C++ are specified with the pragma preprocessing directive. The syntax of an OpenMP directive is formally specified by the grammar in Appendix B, and informally as follows: 5 6 7 #pragma omp directive-name [clause[ [,] clause]...] new-line 8 9 10 11 12 Each directive starts with #pragma omp. The remainder of the directive follows the conventions of the C and C++ standards for compiler directives. In particular, white space can be used before and after the #, and sometimes white space must be used to separate the words in a directive. Preprocessing tokens following the #pragma omp are subject to macro replacement. 13 14 Some OpenMP directives may be composed of consecutive #pragma preprocessing directives if specified in their syntax. 15 Directives are case-sensitive. 16 17 An OpenMP executable directive applies to at most one succeeding statement, which must be a structured block. C/C++ Fortran OpenMP directives for Fortran are specified as follows: 18 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 27 and Section 2.1.2 on page 28. 19 20 21 22 26 OpenMP API • Version 4.0 - July 2013 1 2 Directives are case insensitive. Directives cannot be embedded within continued statements, and statements cannot be embedded within directives. 3 4 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. Fortran 5 6 7 8 Only one directive-name can be specified per directive (note that this includes combined directives, see Section 2.10 on page 95). 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. 9 10 11 12 Some data-sharing attribute clauses (Section 2.14.3 on page 155), data copying clauses (Section 2.14.4 on page 173), the threadprivate directive (Section 2.14.2 on page 150) and the flush directive (Section 2.12.7 on page 134) accept a list. A list consists of a comma-separated collection of one or more list items. 13 14 15 A list item is a variable or array section, subject to the restrictions specified in Section 2.4 on page 42 and in each of the sections describing clauses and directives for which a list appears. C/C++ C/C++ Fortran A list item is a variable, array section or common block name (enclosed in slashes), subject to the restrictions specified in Section 2.4 on page 42 and in each of the sections describing clauses and directives for which a list appears. 16 17 18 Fortran 19 Fortran 20 21 2.1.1 Fixed Source Form Directives The following sentinels are recognized in fixed form source files: !$omp | c$omp | *$omp 22 23 24 25 26 Sentinels must start in column 1 and appear as a single word with no intervening characters. Fortran fixed form line length, white space, continuation, and column rules apply to the directive line. Initial directive lines must have a space or zero in column 6, and continuation directive lines must have a character other than a space or a zero in column 6. Chapter 2 Directives 27 Fortran (cont.) Comments may appear on the same line as a directive. The exclamation point initiates a comment when it appears after column 6. The comment extends to the end of the source line and is ignored. If the first non-blank character after the directive sentinel of an initial or continuation directive line is an exclamation point, the line is ignored. 1 2 3 4 5 6 7 Note – in the following example, the three formats for specifying the directive are equivalent (the first line represents the position of the first 9 columns): 8 c23456789 9 !$omp parallel do shared(a,b,c) 10 11 c$omp parallel do 12 c$omp+shared(a,b,c) 13 14 15 c$omp paralleldoshared(a,b,c) 2.1.2 Free Source Form Directives The following sentinel is recognized in free form source files: 16 !$omp 17 18 19 20 21 22 23 24 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 non-blank 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. 25 26 27 28 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 non-blank character after the directive sentinel is an exclamation point, the line is ignored. 28 OpenMP API • Version 4.0 - July 2013 Fortran (cont.) 1 2 3 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 set of keywords: 4 declare reduction 5 declare simd 6 declare target 7 distribute parallel do 8 distribute parallel do simd 9 distribute simd 10 do simd 11 end atomic 12 end critical 13 end distribute 14 end distribute parallel do 15 end distribute parallel do simd 16 end distribute simd 17 end do 18 end do simd 19 end master 20 end ordered 21 end parallel 22 end parallel do 23 end parallel do simd 24 end parallel sections 25 end parallel workshare 26 end sections 27 end simd Chapter 2 Directives 29 Fortran (cont.) 1 end single 2 end target 3 end target data 4 end target teams 5 end target teams distribute 6 end target teams distribute parallel do 7 end target teams distribute parallel do simd 8 end target teams distribute simd 9 end task 10 end task group 11 end teams 12 end teams distribute 13 end teams distribute parallel do 14 end teams distribute parallel do simd 15 end teams distribute simd 16 end workshare 17 parallel do 18 parallel do simd 19 parallel sections 20 parallel workshare 21 target data 22 target teams 23 target teams distribute 24 target teams distribute parallel do 25 target teams distribute parallel do simd 26 target teams distribute simd 30 OpenMP API • Version 4.0 - July 2013 1 target update 2 teams distribute 3 teams distribute parallel do 4 teams distribute parallel do simd 5 teams distribute simd 6 7 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): 8 !23456789 9 !$omp parallel do & 10 !$omp shared(a,b,c) 11 12 !$omp parallel & 13 !$omp&do shared(a,b,c) 14 15 !$omp paralleldo shared(a,b,c) 16 Fortran 17 2.1.3 Stand-Alone Directives 18 Summary 19 Stand-alone directives are executable directives that have no associated user code. 20 Description 21 22 23 24 25 Stand-alone directives do not have any associated executable user code. Instead, they represent executable statements that typically do not have succinct equivalent statements in the base languages. There are some restrictions on the placement of a stand-alone directive within a program. A stand-alone directive may be placed only at a point where a base language executable statement is allowed. Chapter 2 Directives 31 1 Restrictions 2 3 For C/C++, a stand-alone directive may not be used in place of the statement following an if, while, do, switch, or label. See Appendix B for the formal grammar. C/C++ C/C++ Fortran For Fortran, a stand-alone directive may not be used as the action statement in an if statement or as the executable statement following a label if the label is referenced in the program. 4 5 6 Fortran 7 2.2 Conditional Compilation 8 9 10 In implementations that support a preprocessor, the _OPENMP macro name is defined to have the decimal value yyyymm where yyyy and mm are the year and month designations of the version of the OpenMP API that the implementation supports. 11 12 If this macro is the subject of a #define or a #undef preprocessing directive, the behavior is unspecified. Fortran The OpenMP API requires Fortran lines to be compiled conditionally, as described in the following sections. 13 14 16 Fixed Source Form Conditional Compilation Sentinels 17 18 The following conditional compilation sentinels are recognized in fixed form source files: 15 2.2.1 !$ | *$ | c$ 19 20 To enable conditional compilation, a line with a conditional compilation sentinel must satisfy the following criteria: 21 22 • The sentinel must start in column 1 and appear as a single word with no intervening white space. 32 OpenMP API • Version 4.0 - July 2013 Fortran (cont.) 1 2 • After the sentinel is replaced with two spaces, initial lines must have a space or zero 3 4 5 • After the sentinel is replaced with two spaces, continuation lines must have a 6 7 If these criteria are met, the sentinel is replaced by two spaces. If these criteria are not met, the line is left unchanged. 8 9 10 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): 11 c23456789 12 !$ 10 iam = omp_get_thread_num() + 13 !$ in column 6 and only white space and numbers in columns 1 through 5. character other than a space or zero in column 6 and only white space in columns 1 through 5. & index 14 15 16 #ifdef _OPENMP 17 10 iam = omp_get_thread_num() + 18 & 19 #endif 21 Free Source Form Conditional Compilation Sentinel 22 The following conditional compilation sentinel is recognized in free form source files: 20 2.2.2 index !$ 23 24 To enable conditional compilation, a line with a conditional compilation sentinel must satisfy the following criteria: 25 • The sentinel can appear in any column but must be preceded only by white space. 26 • The sentinel must appear as a single word with no intervening white space. 27 • Initial lines must have a space after the sentinel. Chapter 2 Directives 33 1 2 3 4 • Continued lines must have an ampersand as the last non-blank character on the line, 5 6 If these criteria are met, the sentinel is replaced by two spaces. If these criteria are not met, the line is left unchanged. 7 8 9 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): 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. 10 c23456789 11 !$ iam = omp_get_thread_num() + 12 !$& & index 13 14 #ifdef _OPENMP 15 iam = omp_get_thread_num() + 16 & index 17 #endif 18 Fortran 19 2.3 Internal Control Variables 20 21 22 23 24 25 26 27 An OpenMP implementation must act as if there are 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. 28 29 30 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 36. 34 OpenMP API • Version 4.0 - July 2013 1 2.3.1 ICV Descriptions 2 The following ICVs store values that affect the operation of parallel regions. 3 4 5 • dyn-var - controls whether dynamic adjustment of the number of threads is enabled 6 7 • nest-var - controls whether nested parallelism is enabled for encountered parallel 8 9 • nthreads-var - controls the number of threads requested for encountered parallel for encountered parallel regions. There is one copy of this ICV per data environment. regions. There is one copy of this ICV per data environment. regions. There is one copy of this ICV per data environment. 10 11 • thread-limit-var - controls the maximum number of threads participating in the 12 13 • max-active-levels-var - controls the maximum number of nested active parallel 14 15 16 • place-partition-var – controls the place partition available to the execution 17 18 19 • active-levels-var - the number of nested, active parallel regions enclosing the current 20 21 22 • levels-var - the number of nested parallel regions enclosing the current task such that 23 24 25 26 • bind-var - controls the binding of OpenMP threads to places. When binding is 27 The following ICVs store values that affect the operation of loop regions. 28 29 • run-sched-var - controls the schedule that the runtime schedule clause uses for 30 31 • def-sched-var - controls the implementation defined default scheduling of loop 32 The following ICVs store values that affect the program execution. 33 34 • stacksize-var - controls the stack size for threads that the OpenMP implementation 35 36 • wait-policy-var - controls the desired behavior of waiting threads. There is one copy 37 38 • cancel-var - controls the desired behavior of the cancel construct and cancellation contention group. There is one copy of this ICV per data environment. regions. There is one copy of this ICV per device. environment for encountered parallel regions. There is one copy of this ICV per implicit task. task such that all of the parallel regions are enclosed by the outermost initial task region on the current device. There is one copy of this ICV per data environment. all of the parallel regions are enclosed by the outermost initial task region on the current device. There is one copy of this ICV per data environment. requested, the variable indicates that the execution environment is advised not to move threads between places. The variable can also provide default thread affinity policies. There is one copy of this ICV per data environment. loop regions. There is one copy of this ICV per data environment. regions. There is one copy of this ICV per device. creates. There is one copy of this ICV per device. of this ICV per device. points. There is one copy of the ICV for the whole program (the scope is global). Chapter 2 Directives 35 • default-device-var - controls the default target device. There is one copy of this ICV 1 2 3 per data environment. 2.3.2 ICV Initialization The following table shows the ICVs, associated environment variables, and initial values: 4 5 ICV Environment Variable Initial value dyn-var OMP_DYNAMIC See comments below nest-var OMP_NESTED false nthreads-var OMP_NUM_THREADS Implementation defined run-sched-var OMP_SCHEDULE Implementation defined def-sched-var (none) Implementation defined bind-var OMP_PROC_BIND Implementation defined stacksize-var OMP_STACKSIZE Implementation defined wait-policy-var OMP_WAIT_POLICY Implementation defined thread-limit-var OMP_THREAD_LIMIT Implementation defined max-active-levels-var OMP_MAX_ACTIVE_LEVELS See comments below active-levels-var (none) zero levels-var (none) zero place-partition-var OMP_PLACES Implementation defined cancel-var OMP_CANCELLATION false default-device-var OMP_DEFAULT_DEVICE Implementation defined 6 Comments: 7 • Each device has its own ICVs. 8 • The value of the nthreads-var ICV is a list. 9 • The value of the bind-var ICV is a list. 10 11 • The initial value of dyn-var is implementation defined if the implementation supports 12 13 14 • The initial value of max-active-levels-var is the number of levels of parallelism that dynamic adjustment of the number of threads; otherwise, the initial value is false. the implementation supports. See the definition of supporting n levels of parallelism in Section 1.2.6 on page 12 for further details. 36 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 The host and target device ICVs are initialized before any OpenMP API construct or OpenMP API routine executes. After the initial values are assigned, the values of any OpenMP environment variables that were set by the user are read and the associated ICVs for the host device are modified accordingly. The method for initializing a target device's ICVs is implementation defined. 6 Cross References: 7 • OMP_SCHEDULE environment variable, see Section 4.1 on page 238. 8 • OMP_NUM_THREADS environment variable, see Section 4.2 on page 239. 9 • OMP_DYNAMIC environment variable, see Section 4.3 on page 240. 10 • OMP_PROC_BIND environment variable, see Section 4.4 on page 241. 11 • OMP_PLACES environment variable, see Section 4.5 on page 241. 12 • OMP_NESTED environment variable, see Section 4.6 on page 243. 13 • OMP_STACKSIZE environment variable, see Section 4.7 on page 244. 14 • OMP_WAIT_POLICY environment variable, see Section 4.8 on page 245. 15 • OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.9 on page 245. 16 • OMP_THREAD_LIMIT environment variable, see Section 4.10 on page 246. 17 • OMP_CANCELLATION environment variable, see Section 4.11 on page 246. 18 • OMP_DEFAULT_DEVICE environment variable, see Section 4.13 on page 248. 19 20 21 2.3.3 Modifying and Retrieving ICV Values The following table shows the method for modifying and retrieving the values of ICVs through OpenMP API routines: ICV Ways to modify value Way to retrieve value dyn-var omp_set_dynamic() omp_get_dynamic() nest-var omp_set_nested() omp_get_nested() nthreads-var omp_set_num_threads() omp_get_max_threads() run-sched-var omp_set_schedule() omp_get_schedule() def-sched-var (none) (none) bind-var (none) omp_get_proc_bind() stacksize-var (none) (none) wait-policy-var (none) (none) thread-limit-var thread_limit clause omp_get_thread_limit() Chapter 2 Directives 37 ICV Ways to modify value Way to retrieve value max-active-levels-var omp_set_max_active_levels() omp_get_max_active_levels() active-levels-var (none) omp_get_active_levels() levels-var (none) omp_get_level() place-partition-var (none) (none) cancel-var (none) omp_get_cancellation() default-device-var omp_set_default_device() omp_get_default_device() 1 Comments: 2 3 4 • The value of the nthreads-var ICV is a list. The runtime call 5 6 • The value of the bind-var ICV is a list. The runtime call omp_get_proc_bind() 7 Cross References: 8 • thread_limit clause of the teams construct, see Section 2.9.5 on page 86. 9 • omp_set_num_threads routine, see Section 3.2.1 on page 189. 10 • omp_get_max_threads routine, see Section 3.2.3 on page 192. 11 • omp_set_dynamic routine, see Section 3.2.7 on page 197. 12 • omp_get_dynamic routine, see Section 3.2.8 on page 198. 13 • omp_get_cancellation routine, see Section 3.2.9 on page 199. 14 • omp_set_nested routine, see Section 3.2.10 on page 200. 15 • omp_get_nested routine, see Section 3.2.11 on page 201. 16 • omp_set_schedule routine, see Section 3.2.12 on page 203. 17 • omp_get_schedule routine, see Section 3.2.13 on page 205. 18 • omp_get_thread_limit routine, see Section 3.2.14 on page 206. 19 • omp_set_max_active_levels routine, see Section 3.2.15 on page 207. 20 • omp_get_max_active_levels routine, see Section 3.2.16 on page 209. 21 • omp_get_level routine, see Section 3.2.17 on page 210. 22 • omp_get_active_level routine, see Section 3.2.20 on page 214. 23 • omp_get_proc_bind routine, see Section 3.2.22 on page 216 24 • omp_set_default_device routine, see Section 3.2.23 on page 218. 25 • omp_get_default_device routine, see Section 3.2.24 on page 219. omp_set_num_threads() sets the value of the first element of this list, and omp_get_max_threads() retrieves the value of the first element of this list. retrieves the value of the first element of this list. 38 OpenMP API • Version 4.0 - July 2013 1 2.3.4 How ICVs are Scoped The following table shows the ICVs and their scope:: 2 ICV Scope dyn-var data environment nest-var data environment nthreads-var data environment run-sched-var data environment def-sched-var device bind-var data environment stacksize-var device wait-policy-var device thread-limit-var data environment max-active-levels-var device active-levels-var data environment levels-var data environment place-partition-var implicit task cancel-var device default-device-var data environment 3 Comments: 4 • There is one copy per device of each ICV with device scope 5 • Each data environment has its own copies of ICVs with data environment scope 6 • Each implicit task has its own copy of ICVs with implicit task scope 7 8 Calls to OpenMP API routines retrieve or modify data environment scoped ICVs in the data environment of their binding tasks. 9 2.3.4.1 How the Per-Data Environment ICVs Work 10 11 12 When a task construct or parallel construct is encountered, the generated task(s) inherit the values of the data environment scoped ICVs from the generating task's ICV values. 13 14 15 When a task construct is encountered, the generated task inherits the value of nthreads-var from the generating task's nthreads-var value. When a parallel construct is encountered, and the generating task's nthreads-var list contains a single Chapter 2 Directives 39 1 2 3 4 5 element, the generated task(s) inherit that list as the value of nthreads-var. When a parallel construct is encountered, and the generating task's nthreads-var list contains multiple elements, the generated task(s) inherit the value of nthreads-var as the list obtained by deletion of the first element from the generating task's nthreads-var value. The bind-var ICV is handled in the same way as the nthreads-var ICV. 6 7 8 9 10 11 When a device construct is encountered, the new device data environment inherits the values of the data environment scoped ICVs from the enclosing device data environment of the device that will execute the region. If a teams construct with a thread_limit clause is encountered, the thread-limit-var ICV of the new device data environment is not inherited but instead is set to a value that is less than or equal to the value specified in the clause. 12 13 14 When encountering a loop worksharing region with schedule(runtime), all implicit task regions that constitute the binding parallel region must have the same value for run-sched-var in their data environments. Otherwise, the behavior is unspecified. 15 2.3.5 ICV Override Relationships The override relationships among construct clauses and ICVs are shown in the following table: 16 17 40 ICV construct clause, if used dyn-var (none) nest-var (none) nthreads-var num_threads run-sched-var schedule def-sched-var schedule bind-var proc_bind stacksize-var (none) wait-policy-var (none) thread-limit-var (none) max-active-levels-var (none) active-levels-var (none) levels-var (none) place-partition-var (none) cancel-var (none) default-device-var (none) OpenMP API • Version 4.0 - July 2013 1 Comments: 2 3 • The num_threads clause overrides the value of the first element of the 4 5 • If bind-var is not set to false then the proc_bind clause overrides the value of the 6 Cross References: 7 • parallel construct, see Section 2.5 on page 44. 8 • proc_bind clause, Section 2.5 on page 44. 9 • num_threads clause, see Section 2.5.1 on page 47. nthreads-var ICV. first elements of the bind-var ICV; otherwise, the proc_bind clause has no effect. 10 • Loop construct, see Section 2.7.1 on page 53. 11 • schedule clause, see Section 2.7.1.1 on page 59. Chapter 2 Directives 41 1 2.4 Array Sections 2 3 An array section designates a subset of the elements in an array. An array section can appear only in clauses where it is explicitly allowed. 4 5 To specify an array section in an OpenMP construct, array subscript expressions are extended with the following syntax: 6 [ lower-bound : length ] or 7 [ lower-bound : ] or 8 [ : length ] or 9 [ : ] C/C++ 10 The array section must be a subset of the original array. 11 12 13 Array sections are allowed on multidimensional arrays. Base language array subscript expressions can be used to specify length-one dimensions of multidimensional array sections. 14 15 The lower-bound and length are integral type expressions. When evaluated they represent a set of integer values as follows: 16 { lower-bound, lower-bound + 1, lower-bound + 2,... , lower-bound + length - 1 } 17 The lower-bound and length must evaluate to non-negative integers. 18 19 When the size of the array dimension is not known, the length must be specified explicitly. 20 21 When the length is absent, it defaults to the size of the array dimension minus the lower-bound. 22 When the lower-bound is absent it defaults to 0. 42 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 Note – The following are examples of array sections: a[0:6] a[:6] a[1:10] a[1:] b[10][:][:0] c[1:10][42][0:6] The first two examples are equivalent. If a is declared to be an eleven element array, the third and fourth examples are equivalent. The fifth example is a zero-length array section. The last example is not contiguous. C/C++ 12 Fortran 13 14 Fortran has built-in support for array sections but the following restrictions apply for OpenMP constructs: 15 • A stride expression may not be specified. 16 17 • The upper bound for the last dimension of an assumed-size dummy array must be specified. Fortran 18 Restrictions 19 Restrictions to array sections are as follows: 20 • An array section can appear only in clauses where it is explicitly allowed. 21 • An array section can only be specified for a base language identifier. 22 23 • The type of the variable appearing in an array section must be array, pointer, C/C++ reference to array, or reference to pointer. C/C++ C++ 24 • An array section cannot be used in a C++ user-defined []-operator. C++ Chapter 2 Directives 43 1 2.5 parallel Construct 2 Summary 3 4 This fundamental construct starts parallel execution. See Section 1.3 on page 14 for a general description of the OpenMP execution model. 5 Syntax 6 The syntax of the parallel construct is as follows: C/C++ #pragma omp parallel [clause[ [, ]clause] ...] new-line structured-block where clause is one of the following: 7 if(scalar-expression) num_threads(integer-expression) default(shared | none) private(list) firstprivate(list) shared(list) copyin(list) reduction(redution-identifier :list) proc_bind(master | close | spread) C/C++ 8 44 OpenMP API • Version 4.0 - July 2013 Fortran 1 The syntax of the parallel construct is as follows: !$omp parallel [clause[[,] clause]...] structured-block !$omp end parallel 2 where clause is one of the following: if(scalar-logical-expression) num_threads(scalar-integer-expression) default(private | firstprivate | shared | none) private(list) firstprivate(list) shared(list) copyin(list) reduction(reduction-identifier : list) proc_bind(master | close | spread) 3 The end parallel directive denotes the end of the parallel construct. Fortran 4 Binding 5 6 The binding thread set for a parallel region is the encountering thread. The encountering thread becomes the master thread of the new team. 7 Description 8 9 10 11 12 13 14 15 When a thread encounters a parallel construct, a team of threads is created to execute the parallel region (see Section 2.5.1 on page 47 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. Chapter 2 Directives 45 1 2 3 4 The optional proc_bind clause, described in Section 2.5.2 on page 49, specifies the mapping of OpenMP threads to places within the current place partition, that is, within the places listed in the place-partition-var ICV for the implicit task of the encountering thread. 5 6 7 8 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. 9 10 11 12 13 14 15 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. 16 17 18 19 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 2.11 on page 113). 20 21 22 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. 23 24 25 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 2.5.1 on page 47, and it becomes the master of that new team. 26 27 28 29 30 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. 31 Restrictions 32 Restrictions to the parallel construct are as follows: 33 • A program that branches into or out of a parallel region is non-conforming. 34 35 • A program must not depend on any ordering of the evaluations of the clauses of the 36 • At most one if clause can appear on the directive. 37 • At most one proc_bind clause can appear on the directive. parallel directive, or on any side effects of the evaluations of the clauses. 46 OpenMP API • Version 4.0 - July 2013 1 2 • At most one num_threads clause can appear on the directive. The num_threads 3 4 5 • A throw executed inside a parallel region must cause execution to resume expression must evaluate to a positive integer value. C/C++ within the same parallel region, and the same thread that threw the exception must catch it. C/C++ Fortran • Unsynchronized use of Fortran I/O statements by multiple threads on the same unit 6 7 has unspecified behavior. Fortran Cross References 8 9 10 • default, shared, private, firstprivate, and reduction clauses, see 11 • copyin clause, see Section 2.14.4 on page 173. 12 • omp_get_thread_num routine, see Section 3.2.4 on page 193. Section 2.14.3 on page 155. 14 Determining the Number of Threads for a parallel Region 15 16 17 18 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-levels-var, and nest-var ICVs are used to determine the number of threads to use in the region. 19 20 21 22 23 24 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. 13 2.5.1 Chapter 2 Directives 47 When a thread encounters a parallel construct, the number of threads is determined according to Algorithm 2.1. 1 2 Algorithm 2.1 let ThreadsBusy be the number of OpenMP threads currently executing in this contention group; let ActiveParRegions be the number of enclosing active parallel regions; if an if clause exists then let IfClauseValue be the value of the if clause expression; else let IfClauseValue = true; if a num_threads clause exists then let ThreadsRequested be the value of the num_threads clause expression; else let ThreadsRequested = value of the first element of nthreads-var; let ThreadsAvailable = (thread-limit-var - ThreadsBusy + 1); if (IfClauseValue = false) then number of threads = 1; else if (ActiveParRegions >= 1) and (nest-var = false) then number of threads = 1; else if (ActiveParRegions = max-active-levels-var) then number of threads = 1; else if (dyn-var = true) and (ThreadsRequested <= ThreadsAvailable) then number of threads = [ 1 : ThreadsRequested ]; else if (dyn-var = true) and (ThreadsRequested > ThreadsAvailable) then number of threads = [ 1 : ThreadsAvailable ]; else if (dyn-var = false) and (ThreadsRequested <= ThreadsAvailable) then number of threads = ThreadsRequested; else if (dyn-var = false) and (ThreadsRequested > ThreadsAvailable) then behavior is implementation defined; 48 OpenMP API • Version 4.0 - July 2013 1 2 3 Note – Since the initial value of the dyn-var ICV is implementation defined, programs that depend on a specific number of threads for correct execution should explicitly disable dynamic adjustment of the number of threads. 4 Cross References 5 6 • nthreads-var, dyn-var, thread-limit-var, max-active-levels-var, and nest-var ICVs, see 7 Section 2.3 on page 34. 2.5.2 Controlling OpenMP Thread Affinity 8 9 10 11 12 When creating a team for a parallel region, the proc_bind clause specifies a policy for assigning OpenMP threads to places within the current place partition, that is, the places listed in the place-partition-var ICV for the implicit task of the encountering thread. Once a thread is assigned to a place, the OpenMP implementation should not move it to another place. 13 14 15 16 The master thread affinity policy instructs the execution environment to assign every thread in the team to the same place as the master thread. The place partition is not changed by this policy, and each implicit task inherits the place-partition-var ICV of the parent implicit task. 17 18 19 20 21 22 23 The close thread affinity policy instructs the execution environment to assign the threads to places close to the place of the parent thread. The master thread executes on the parent’s place and the remaining threads in the team execute on places from the place list consecutive from the parent’s position in the list, with wrap around with respect to the place partition of the parent thread’s implicit task. The place partition is not changed by this policy, and each implicit task inherits the place-partition-var ICV of the parent implicit task. 24 25 26 27 28 29 30 31 The purpose of the spread thread affinity policy is to create a sparse distribution for a team of T threads among the P places of the parent's place partition. It accomplishes this by first subdividing the parent partition into T subpartitions if T is less than or equal to P, or P subpartitions if T is greater than P. Then it assigns one thread (T<=P) or a set of threads (T>P) to each subpartition. The place-partition-var ICV of each implicit task is set to its subpartition. The subpartitioning is not only a mechanism for achieving a sparse distribution, it also defines a subset of places for a thread to use when creating a nested parallel region. Chapter 2 Directives 49 • T<=P. The parent's partition is split into T subpartitions, where each subpartition 1 2 3 4 5 6 7 contains at least S=floor(P/T) consecutive places. A single thread is assigned to each subpartition. The master thread executes on the place of the parent thread and is assigned to the subpartition that includes that place. For the other threads, assignment is to the first place in the corresponding subpartition. When T does not divide P evenly, the assignment of the remaining P-T*S places into subpartitions is implementation defined. 8 9 10 • T>P. The parent's partition is split into P unit-sized subpartitions. Each place is 11 12 13 14 For the close and spread thread affinity policies, the threads with the smallest thread numbers execute on the place of the master thread, then the threads with the next smaller thread numbers execute on the next place in the partition; and so on, with wrap around with respect to the encountering thread's place partition. 15 16 17 The determination of whether the affinity request can be fulfilled is implementation defined. If the affinity request cannot be fulfilled, then the number of threads in the team and their mapping to places are implementation defined. assigned S=floor(T/P) threads. When P does not divide T evenly, the assignment of the remaining T-P*S threads into places is implementation defined. 50 OpenMP API • Version 4.0 - July 2013 1 2.6 Canonical Loop Form C/C++ A loop has canonical loop form if it conforms to the following: 2 for (init-expr; test-expr; incr-expr) structured-block init-expr One of the following: var = lb integer-type var = lb random-access-iterator-type var = lb pointer-type var = lb test-expr One of the following: var relational-op b b relational-op var incr-expr One of the following: ++var var++ --var var-var += incr var -= incr var = var + incr var = incr + var var = var - incr var One of the following: A variable of a signed or unsigned integer type. For C++, a variable of a random access iterator type. For C, a variable of a pointer type. If this variable would otherwise be shared, it is implicitly made private in the loop construct. This variable must not be modified during the execution of the for-loop other than in incr-expr. Unless the variable is specified lastprivate on the loop construct, its value after the loop is unspecified. relational-op One of the following: < <= > >= lb and b Loop invariant expressions of a type compatible with the type of var. incr A loop invariant integer expression. Chapter 2 Directives 51 1 2 3 The canonical form allows the iteration count of all associated loops to be computed before executing the outermost loop. The computation is performed for each loop in an integer type. This type is derived from the type of var as follows: 4 • If var is of an integer type, then the type is the type of var. 5 6 • For C++, if var is of a random access iterator type, then the type is the type that 7 • For C, if var is of a pointer type, then the type is ptrdiff_t. 8 9 The behavior is unspecified if any intermediate result required to compute the iteration count cannot be represented in the type determined above. 10 11 12 There is no implied synchronization during the evaluation of the lb, b, or incr expressions. It is unspecified whether, in what order, or how many times any side effects within the lb, b, or incr expressions occur. 13 14 15 16 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. Restrictions 18 The following restrictions also apply: 19 20 21 22 • If test-expr is of the form var relational-op b and relational-op is < or <= then 23 24 25 26 • If test-expr is of the form b relational-op var and relational-op is < or <= then 27 28 • For C++, in the simd construct the only random access iterator types that are would be used by std::distance applied to variables of the type of var. incr-expr must cause var to increase on each iteration of the loop. If test-expr is of the form var relational-op b and relational-op is > or >= then incr-expr must cause var to decrease on each iteration of the loop. incr-expr must cause var to decrease on each iteration of the loop. If test-expr is of the form b relational-op var and relational-op is > or >= then incr-expr must cause var to increase on each iteration of the loop. allowed for var are pointer types. C/C++ 29 52 OpenMP API • Version 4.0 - July 2013 1 2.7 Worksharing Constructs A worksharing construct distributes the execution of the associated region among the members of the team that encounters it. Threads execute portions of the region in the context of the implicit tasks each one is executing. If the team consists of only one thread then the worksharing region is not executed in parallel. 2 3 4 5 6 7 8 9 10 11 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. 12 13 The OpenMP API defines the following worksharing constructs, and these are described in the sections that follow: 14 • loop construct 15 • sections construct 16 • single construct 17 • workshare construct 18 Restrictions 19 The following restrictions apply to worksharing constructs: 20 21 22 • Each worksharing region must be encountered by all threads in a team or by none at 23 24 • The sequence of worksharing regions and barrier regions encountered must be the 25 all, unless cancellation has been requested for the innermost enclosing parallel region. same for every thread in a team. 2.7.1 Loop Construct 26 Summary 27 28 29 30 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. Chapter 2 Directives 53 1 Syntax 2 The syntax of the loop construct is as follows: C/C++ #pragma omp for [clause[[,] clause] ... ] new-line for-loops where clause is one of the following: 3 private(list) firstprivate(list) lastprivate(list) reduction(reduction-identifier: list) schedule(kind[, chunk_size]) collapse(n) ordered nowait The for directive places restrictions on the structure of all associated for-loops. Specifically, all associated for-loops must have canonical loop form (see Section 2.6 on page 51). 4 5 6 C/C++ Fortran The syntax of the loop construct is as follows: 7 !$omp do [clause[[,] clause] ... ] do-loops [!$omp end do [nowait] ] where clause is one of the following: 8 private(list) firstprivate(list) lastprivate(list) reduction({reduction-identifier:list) 54 OpenMP API • Version 4.0 - July 2013 schedule(kind[, chunk_size]) collapse(n) ordered 1 2 If an end do directive is not specified, an end do directive is assumed at the end of the do-loop. 3 4 5 6 All associated do-loops must be do-constructs as defined by the Fortran standard. If an end do directive follows a do-construct in which several loop statements share a DO termination statement, then the directive can only be specified for the outermost of these DO statements. 7 8 9 If any of the loop iteration variables would otherwise be shared, they are implicitly made private on the loop construct. Unless the loop iteration variables are specified lastprivate on the loop construct, their values after the loop are unspecified. Fortran 10 Binding 11 12 13 14 The binding thread set for a loop region is the current team. A loop region binds to the innermost enclosing parallel region. Only the threads of the team executing the binding parallel region participate in the execution of the loop iterations and the implied barrier of the loop region if the barrier is not eliminated by a nowait clause. 15 Description 16 17 The loop construct is associated with a loop nest consisting of one or more loops that follow the directive. 18 19 There is an implicit barrier at the end of a loop construct unless a nowait clause is specified. 20 21 22 23 The collapse clause may be used to specify how many loops are associated with the loop construct. The parameter of the collapse clause must be a constant positive integer expression. If no collapse clause is present, the only loop that is associated with the loop construct is the one that immediately follows the loop directive. 24 25 26 27 If more than one loop is associated with the loop construct, then the iterations of all associated loops are collapsed into one larger iteration space that is then divided according to the schedule clause. The sequential execution of the iterations in all associated loops determines the order of the iterations in the collapsed iteration space. Chapter 2 Directives 55 1 2 3 The iteration count for each associated loop is computed before entry to the outermost loop. If execution of any associated loop changes any of the values used to compute any of the iteration counts, then the behavior is unspecified. 4 5 The integer type (or kind, for Fortran) used to compute the iteration count for the collapsed loop is implementation defined. 6 7 8 9 10 11 12 13 14 15 16 A worksharing loop has logical iterations numbered 0,1,...,N-1 where N is the number of loop iterations, and the logical numbering denotes the sequence in which the iterations would be executed if the associated loop(s) were executed by a single thread. The schedule clause specifies how iterations of the associated loops are divided into contiguous non-empty subsets, called chunks, and how these chunks are distributed among threads of the team. Each thread executes its assigned chunk(s) in the context of its implicit task. The chunk_size expression is evaluated using the original list items of any variables that are made private in the loop construct. It is unspecified whether, in what order, or how many times, any side effects of the evaluation of this expression occur. The use of a variable in a schedule clause expression of a loop construct causes an implicit reference to the variable in all enclosing constructs. 17 18 19 20 21 Different loop regions with the same schedule and iteration count, even if they occur in the same parallel region, can distribute iterations among threads differently. The only exception is for the static schedule as specified in Table 2-1. Programs that depend on which thread executes a particular iteration under any other circumstances are non-conforming. 22 23 See Section 2.7.1.1 on page 59 for details of how the schedule for a worksharing loop is determined. 24 The schedule kind can be one of those specified in Table 2-1. 25 56 OpenMP API • Version 4.0 - July 2013 1 TABLE 2-1 static schedule clause kind values When schedule(static, chunk_size) is specified, iterations are divided into chunks of size chunk_size, and the chunks are assigned to the threads in the team in a round-robin fashion in the order of the thread number. When no chunk_size is specified, the iteration space is divided into chunks that are approximately equal in size, and at most one chunk is distributed to each thread. Note that the size of the chunks is unspecified in this case. A compliant implementation of the static schedule must ensure that the same assignment of logical iteration numbers to threads will be used in two loop regions if the following conditions are satisfied: 1) both loop regions have the same number of loop iterations, 2) both loop regions have the same value of chunk_size specified, or both loop regions have no chunk_size specified, 3) both loop regions bind to the same parallel region, and 4) neither loop is associated with a SIMD construct. A data dependence between the same logical iterations in two such loops is guaranteed to be satisfied allowing safe use of the nowait clause. dynamic When schedule(dynamic, chunk_size) is specified, the iterations are distributed to threads in the team in chunks as the threads request them. Each thread executes a chunk of iterations, then requests another chunk, until no chunks remain to be distributed. Each chunk contains chunk_size iterations, except for the last chunk to be distributed, which may have fewer iterations. When no chunk_size is specified, it defaults to 1. guided When schedule(guided, chunk_size) is specified, the iterations are assigned to threads in the team in chunks as the executing threads request them. Each thread executes a chunk of iterations, then requests another chunk, until no chunks remain to be assigned. For a chunk_size of 1, the size of each chunk is proportional to the number of unassigned iterations divided by the number of threads in the team, decreasing to 1. For a chunk_size with value k (greater than 1), the size of each chunk is determined in the same way, with the restriction that the chunks do not contain fewer than k iterations (except for the last chunk to be assigned, which may have fewer than k iterations). When no chunk_size is specified, it defaults to 1. auto 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 57 runtime When schedule(runtime) is specified, the decision regarding scheduling is deferred until run time, and the schedule and chunk size are taken from the run-sched-var ICV. If the ICV is set to auto, the schedule is implementation defined. 1 2 3 4 5 6 Note – For a team of p threads and a loop of n iterations, let n ⁄ p be the integer q that satisfies n = p*q - r, with 0 ≤ r < p . One compliant implementation of the static schedule (with no specified chunk_size) would behave as though chunk_size had been specified with value q. Another compliant implementation would assign q iterations to the first p-r threads, and q-1 iterations to the remaining r threads. This illustrates why a conforming program must not rely on the details of a particular implementation. 7 8 9 10 11 12 A compliant implementation of the guided schedule with a chunk_size value of k would assign q = n ⁄ p iterations to the first available thread and set n to the larger of n-q and p*k. It would then repeat this process until q is greater than or equal to the number of remaining iterations, at which time the remaining iterations form the final chunk. Another compliant implementation could use the same method, except with q = n ⁄ ( 2p ) , and set n to the larger of n-q and 2*p*k. 13 Restrictions 14 Restrictions to the loop construct are as follows: 15 16 • All loops associated with the loop construct must be perfectly nested; that is, there 17 18 • The values of the loop control expressions of the loops associated with the loop 19 • Only one schedule clause can appear on a loop directive. 20 • Only one collapse clause can appear on a loop directive. 21 • chunk_size must be a loop invariant integer expression with a positive value. 22 • The value of the chunk_size expression must be the same for all threads in the team. 23 • The value of the run-sched-var ICV must be the same for all threads in the team. 24 25 • When schedule(runtime) or schedule(auto) is specified, chunk_size must 26 • Only one ordered clause can appear on a loop directive. 27 28 • The ordered clause must be present on the loop construct if any ordered region 29 • The loop iteration variable may not appear in a threadprivate directive. must be no intervening code nor any OpenMP directive between any two loops. construct must be the same for all the threads in the team. not be specified. ever binds to a loop region arising from the loop construct. 58 OpenMP API • Version 4.0 - July 2013 C/C++ 1 • The associated for-loops must be structured blocks. 2 3 • Only an iteration of the innermost associated loop may be curtailed by a continue 4 • No statement can branch to any associated for statement. 5 • Only one nowait clause can appear on a for directive. 6 7 8 • A throw executed inside a loop region must cause execution to resume within the statement. same iteration of the loop region, and the same thread that threw the exception must catch it. C/C++ Fortran • The associated do-loops must be structured blocks. 9 10 11 • Only an iteration of the innermost associated loop may be curtailed by a CYCLE 12 13 • No statement in the associated loops other than the DO statements can cause a branch 14 • The do-loop iteration variable must be of type integer. 15 • The do-loop cannot be a DO WHILE or a DO loop without loop control. statement. out of the loops. Fortran 16 Cross References 17 18 • private, firstprivate, lastprivate, and reduction clauses, see 19 • OMP_SCHEDULE environment variable, see Section 4.1 on page 238. 20 • ordered construct, see Section 2.12.8 on page 138. 21 22 23 24 25 26 27 28 29 30 Section 2.14.3 on page 155. 2.7.1.1 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 34 for details of how the values of the ICVs are determined. If the loop directive does not have a schedule clause then the current value of the def-sched-var ICV determines the schedule. If the loop directive has a schedule clause that specifies the runtime schedule kind then the current value of the run-sched-var ICV determines the schedule. Otherwise, the value of the schedule clause determines the schedule. Figure 2-1 describes how the schedule for a worksharing loop is determined. Chapter 2 Directives 59 1 Cross References 2 • ICVs, see Section 2.3 on page 34. 3 START schedule clause present? No Use def-sched-var schedule kind Yes schedule kind value is runtime? No Use schedule kind specified in schedule clause Yes Use run-sched-var schedule kind 4 5 FIGURE 2-1 2.7.2 Determining the schedule for a worksharing loop. sections Construct Summary 6 The sections construct is a non-iterative 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 8 9 10 60 OpenMP API • Version 4.0 - July 2013 1 Syntax 2 The syntax of the sections construct is as follows: C/C++ #pragma omp sections [clause[[,] clause] ...] new-line { [#pragma omp section new-line] structured-block [#pragma omp section new-line structured-block ] ... } 3 where clause is one of the following: private(list) firstprivate(list) lastprivate(list) reduction(reduction-identifier:list) nowait 4 C/C++ Fortran 5 The syntax of the sections construct is as follows: !$omp sections [clause[[,] clause] ...] [!$omp section] structured-block [!$omp section structured-block ] ... !$omp end sections [nowait] 6 where clause is one of the following: private(list) Chapter 2 Directives 61 firstprivate(list) lastprivate(list) reduction(reduction-identifier:list) 1 Fortran 2 Binding 3 4 5 6 7 The binding thread set for a sections region is the current team. A sections region binds to the innermost enclosing parallel region. Only the threads of the team executing the binding parallel region participate in the execution of the structured blocks and the implied barrier of the sections region if the barrier is not eliminated by a nowait clause. 8 Description 9 10 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. 11 12 The method of scheduling the structured blocks among the threads in the team is implementation defined. 13 14 There is an implicit barrier at the end of a sections construct unless a nowait clause is specified. 15 Restrictions 16 Restrictions to the sections construct are as follows: 17 18 19 • Orphaned section directives are prohibited. That is, the section directives must 20 • The code enclosed in a sections construct must be a structured block. 21 • Only a single nowait clause can appear on a sections directive. appear within the sections construct and must not be encountered elsewhere in the sections region. 62 OpenMP API • Version 4.0 - July 2013 C++ • A throw executed inside a sections region must cause execution to resume within 1 2 3 the same section of the sections region, and the same thread that threw the exception must catch it. C++ 4 Cross References 5 6 • private, firstprivate, lastprivate, and reduction clauses, see 7 8 Section 2.14.3 on page 155. 2.7.3 single Construct Summary 9 10 11 12 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. 13 Syntax 14 The syntax of the single construct is as follows: C/C++ #pragma omp single [clause[[,] clause] ...] new-line structured-block 15 where clause is one of the following: private(list) firstprivate(list) copyprivate(list) nowait 16 C/C++ Chapter 2 Directives 63 Fortran The syntax of the single construct is as follows: 1 !$omp single [clause[[,] clause] ...] structured-block !$omp end single [end_clause[[,] end_clause] ...] where clause is one of the following: 2 private(list) firstprivate(list) and end_clause is one of the following: 3 copyprivate(list) nowait 4 Fortran Binding 5 6 7 8 9 10 The binding thread set for a single region is the current team. A single region binds to the innermost enclosing parallel region. Only the threads of the team executing the binding parallel region participate in the execution of the structured block and the implied barrier of the single region if the barrier is not eliminated by a nowait clause. 11 Description 12 13 14 The method of choosing a thread to execute the structured block is implementation defined. There is an implicit barrier at the end of the single construct unless a nowait clause is specified. 15 Restrictions 16 Restrictions to the single construct are as follows: 17 • The copyprivate clause must not be used with the nowait clause. 18 • At most one nowait clause can appear on a single construct. 64 OpenMP API • Version 4.0 - July 2013 C++ • A throw executed inside a single region must cause execution to resume within the 1 2 same single region, and the same thread that threw the exception must catch it. C++ 3 Cross References 4 • private and firstprivate clauses, see Section 2.14.3 on page 155. 5 • copyprivate clause, see Section 2.14.4.2 on page 175. Fortran 6 7 2.7.4 workshare Construct Summary 8 9 10 The workshare construct divides the execution of the enclosed structured block into separate units of work, and causes the threads of the team to share the work such that each unit is executed only once by one thread, in the context of its implicit task. 11 Syntax 12 The syntax of the workshare construct is as follows: !$omp workshare structured-block !$omp end workshare [nowait] 13 The enclosed structured block must consist of only the following: 14 • array assignments 15 • scalar assignments 16 • FORALL statements 17 • FORALL constructs 18 • WHERE statements 19 • WHERE constructs 20 • atomic constructs Chapter 2 Directives 65 Fortran (cont.) 1 • critical constructs 2 • parallel constructs 3 4 Statements contained in any enclosed critical construct are also subject to these restrictions. Statements in any enclosed parallel construct are not restricted. 5 Binding 6 7 8 9 10 The binding thread set for a workshare region is the current team. A workshare region binds to the innermost enclosing parallel region. Only the threads of the team executing the binding parallel region participate in the execution of the units of work and the implied barrier of the workshare region if the barrier is not eliminated by a nowait clause. 11 Description 12 13 There is an implicit barrier at the end of a workshare construct unless a nowait clause is specified. 14 15 16 17 18 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. 19 The statements in the workshare construct are divided into units of work as follows: 20 21 • For array expressions within each statement, including transformational array intrinsic functions that compute scalar values from arrays: 22 23 • Evaluation of each element of the array expression, including any references to ELEMENTAL functions, is a unit of work. 24 25 • Evaluation of transformational array intrinsic functions may be freely subdivided into any number of units of work. 26 • For an array assignment statement, the assignment of each element is a unit of work. 27 • For a scalar assignment statement, the assignment operation is a unit of work. 28 29 • For a WHERE statement or construct, the evaluation of the mask expression and the 30 31 32 • For a FORALL statement or construct, the evaluation of the mask expression, masked assignments are each a unit of work. expressions occurring in the specification of the iteration space, and the masked assignments are each a unit of work. 66 OpenMP API • Version 4.0 - July 2013 1 2 • For an atomic construct, the atomic operation on the storage location designated as 3 • For a critical construct, the construct is a single unit of work. 4 5 6 • For a parallel construct, the construct is a unit of work with respect to the 7 8 • If none of the rules above apply to a portion of a statement in the structured block, x is a unit of work. workshare construct. The statements contained in the parallel construct are executed by a new thread team. then that portion is a unit of work. 9 10 11 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. 12 13 It is unspecified how the units of work are assigned to the threads executing a workshare region. 14 15 16 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. 17 18 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. 19 20 The workshare directive causes the sharing of work to occur only in the workshare construct, and not in the remainder of the workshare region. 21 Restrictions 22 The following restrictions apply to the workshare construct: 23 24 • All array assignments, scalar assignments, and masked array assignments must be 25 26 • The construct must not contain any user defined function calls unless the function is intrinsic assignments. ELEMENTAL. Fortran Chapter 2 Directives 67 1 2.8 SIMD Constructs 2 2.8.1 simd construct 3 Summary 4 5 6 The simd construct can be applied to a loop to indicate that the loop can be transformed into a SIMD loop (that is, multiple iterations of the loop can be executed concurrently using SIMD instructions). 7 Syntax 8 The syntax of the simd construct is as follows: C/C++ 9 #pragma omp simd [clause[[,] clause] ...] new-line for-loops where clause is one of the following: 10 safelen(length) linear(list[:linear-step]) aligned(list[:alignment]) private(list) lastprivate(list) reduction(reduction-identifier:list) collapse(n) The simd directive places restrictions on the structure of the associated for-loops. Specifically, all associated for-loops must have canonical loop form (Section 2.6 on page 51). 11 12 13 C/C++ 68 OpenMP API • Version 4.0 - July 2013 1 Fortran !$omp simd [clause[[,] clause ...] do-loops [!$omp end simd] 2 where clause is one of the following: safelen(length) linear(list[:linear-step]) aligned(list[:alignment]) private(list) lastprivate(list) reduction(reduction-identifier:list) collapse(n) 3 4 If an end simd directive is not specified, an end simd directive is assumed at the end of the do-loops. 5 6 7 8 All associated do-loops must be do-constructs as defined by the Fortran standard. If an end simd 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. Fortran 9 Binding 10 11 A simd region binds to the current task region. The binding thread set of the simd region is the current team. 12 Description 13 14 The simd construct enables the execution of multiple iterations of the associated loops concurrently by means of SIMD instructions. Chapter 2 Directives 69 1 2 3 4 The collapse clause may be used to specify how many loops are associated with the 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 directive. 5 6 7 8 If more than one loop is associated with the simd construct, then the iterations of all associated loops are collapsed into one larger iteration space that is then executed with SIMD instructions. The sequential execution of the iterations in all associated loops determines the order of the iterations in the collapsed iteration space. 9 10 11 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. 12 13 The integer type (or kind, for Fortran) used to compute the iteration count for the collapsed loop is implementation defined. 14 15 16 17 18 19 20 21 22 A SIMD 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 with no SIMD instructions. If the safelen clause is used then no two iterations executed concurrently with SIMD instructions can have a greater distance in the logical iteration space than its value. The parameter of the safelen clause must be a constant positive integer expression. The number of iterations that are executed concurrently at any given time is implementation defined. Each concurrent iteration will be executed by a different SIMD lane. Each set of concurrent iterations is a SIMD chunk. 23 24 The aligned clause declares that the object to which each list item points is aligned to the number of bytes expressed in the optional parameter of the aligned clause. C/C++ C/C++ Fortran The aligned clause declares that the target of each list item is aligned to the number of bytes expressed in the optional parameter of the aligned clause. 25 26 Fortran 27 28 29 The optional parameter of the aligned clause, alignment, must be a constant positive integer expression. If no optional parameter is specified, implementation-defined default alignments for SIMD instructions on the target platforms are assumed. 30 Restrictions 31 32 • All loops associated with the construct must be perfectly nested; that is, there must be no intervening code nor any OpenMP directive between any two loops. 70 OpenMP API • Version 4.0 - July 2013 1 • The associated loops must be structured blocks. 2 • A program that branches into or out of a simd region is non-conforming. 3 • Only one collapse clause can appear on a simd directive. 4 • A list-item cannot appear in more than one aligned clause. 5 • Only one safelen clause can appear on a simd directive. 6 • No OpenMP construct can appear in the simd region. 7 • The simd region cannot contain calls to the longjmp or setjmp functions. C/C++ C/C++ C 8 • The type of list items appearing in the aligned clause must be array or pointer. C C++ 9 10 • The type of list items appearing in the aligned clause must be array, pointer, 11 • No exception can be raised in the simd region. reference to array, or reference to pointer. C++ Fortran 12 • The do-loop iteration variable must be of type integer. 13 • The do-loop cannot be a DO WHILE or a DO loop without loop control. 14 15 • The type of list items appearing in the aligned clause must be C_PTR or Cray pointer, or the list item must have the POINTER or ALLOCATABLE attribute. Fortran 16 Cross References 17 18 • private, lastprivate, linear and reduction clauses, see Section 2.14.3 on page 155. Chapter 2 Directives 71 1 2.8.2 declare simd construct 2 Summary 3 4 5 6 7 The declare simd construct can be applied to a function (C, C++ and Fortran) or a subroutine (Fortran) to enable the creation of one or more versions that can process multiple arguments using SIMD instructions from a single invocation from a SIMD loop. The declare simd directive is a declarative directive. There may be multiple declare simd directives for a function (C, C++, Fortran) or subroutine (Fortran). 8 Syntax 9 The syntax of the declare simd construct is as follows: C/C++ #pragma omp declare simd [clause[[,] clause] ...] new-line [#pragma omp declare simd [clause[[,] clause] ...] new-line] [...] function definition or declaration where clause is one of the following: 10 simdlen(length) linear(argument-list[:constant-linear-step]) aligned(argument-list[:alignment]) uniform(argument-list) inbranch notinbranch C/C++ 11 Fortran !$omp declare simd( proc-name ) [clause[[,] clause] ...] 12 72 OpenMP API • Version 4.0 - July 2013 1 where clause is one of the following:: simdlen(length) linear(argument-list[:constant-linear-step]) aligned(argument-list[:alignment]) uniform(argument-list) inbranch notinbranch Fortran 2 3 Description 4 5 6 The use of a declare simd construct on a function enables the creation of SIMD versions of the associated function that can be used to process multiple arguments from a single invocation from a SIMD loop concurrently. 7 8 The expressions appearing in the clauses of this directive are evaluated in the scope of the arguments of the function declaration or definition. C/C++ C/C++ Fortran 9 10 11 The use of a declare simd construct enables the creation of SIMD versions of the specified subroutine or function that can be used to process multiple arguments from a single invocation from a SIMD loop concurrently. Fortran 12 13 If a declare simd directive contains multiple SIMD declarations, then one or more SIMD versions will be created for each declaration. 14 15 16 17 18 If a SIMD version is created, the number of concurrent arguments for the function is determined by the simdlen clause. If the simdlen clause is used its value corresponds to the number of concurrent arguments of the function. The parameter of the simdlen clause must be a constant positive integer expression. Otherwise, the number of concurrent arguments for the function is implementation defined. 19 20 The uniform clause declares one or more arguments to have an invariant value for all concurrent invocations of the function in the execution of a single SIMD loop. Chapter 2 Directives 73 C/C++ The aligned clause declares that the object to which each list item points is aligned to the number of bytes expressed in the optional parameter of the aligned clause. 1 2 C/C++ Fortran The aligned clause declares that the target of each list item is aligned to the number of bytes expressed in the optional parameter of the aligned clause. 3 4 Fortran The optional parameter of the aligned clause, alignment, must be a constant positive integer expression. If no optional parameter is specified, implementation-defined default alignments for SIMD instructions on the target platforms are assumed. 5 6 7 8 9 10 11 12 The inbranch clause specifies that the function will always be called from inside a conditional statement of a SIMD loop. The notinbranch clause specifies that the function will never be called from inside a conditional statement of a SIMD loop. If neither clause is specified, then the function may or may not be called from inside a conditional statement of a SIMD loop. 13 Restrictions 14 • Each argument can appear in at most one uniform or linear clause. 15 • At most one simdlen clause can appear in a declare simd directive. 16 • Either inbranch or notinbranch may be specified, but not both. 17 18 • When a constant-linear-step expression is specified in a linear clause it must be a 19 • The function or subroutine body must be a structured block. 20 21 • The execution of the function or subroutine, when called from a SIMD loop, cannot 22 23 • The execution of the function or subroutine cannot have any side effects that would 24 • A program that branches into or out of the function is non-conforming. 25 26 27 • If the function has any declarations, then the declare simd construct for any 28 • The function cannot contain calls to the longjmp or setjmp functions. constant positive integer expression. result in the execution of an OpenMP construct. alter its execution for concurrent iterations of a SIMD chunk. C/C++ declaration that has one must be equivalent to the one specified for the definition. Otherwise, the result is unspecified. C/C++ 74 OpenMP API • Version 4.0 - July 2013 C 1 • The type of list items appearing in the aligned clause must be array or pointer. C C++ 2 • The function cannot contain any calls to throw. 3 4 • The type of list items appearing in the aligned clause must be array, pointer, reference to array, or reference to pointer. C++ Fortran 5 • proc-name must not be a generic name, procedure pointer or entry name. 6 7 • Any declare simd directive must appear in the specification part of a subroutine 8 9 • If a declare simd directive is specified in an interface block for a procedure, it subprogram, function subprogram or interface body to which it applies. must match a declare simd directive in the definition of the procedure. 10 11 • If a procedure is declared via a procedure declaration statement, the procedure 12 13 14 15 • If a declare simd directive is specified for a procedure name with explicit 16 17 • Procedure pointers may not be used to access versions created by the declare 18 19 • The type of list items appearing in the aligned clause must be C_PTR or Cray proc-name should appear in the same specification. interface and a declare simd directive is also specified for the definition of the procedure then the two declare simd directives must match. Otherwise the result is unspecified. simd directive. pointer, or the list item must have the POINTER or ALLOCATABLE attribute. Fortran 20 Cross References 21 • reduction clause, see Section 2.14.3.6 on page 167. 22 • linear clause, see Section 2.14.3.7 on page 172. Chapter 2 Directives 75 1 2.8.3 Loop SIMD construct 2 Summary 3 4 5 The loop SIMD construct specifies a loop that can be executed concurrently using SIMD instructions and that those iterations will also be executed in parallel by threads in the team. 6 Syntax C/C++ #pragma omp for simd [clause[[,] clause] ...] new-line for-loops where clause can be any of the clauses accepted by the for or simd directives with identical meanings and restrictions. 7 8 C/C++ Fortran 9 !$omp do simd [clause[[,] clause] ...] do-loops [!$omp end do simd [nowait]] 10 11 where clause can be any of the clauses accepted by the simd or do directives, with identical meanings and restrictions. 12 13 If an end do simd directive is not specified, an end do simd directive is assumed at the end of the do-loop. Fortran 14 Description 15 16 17 18 19 20 The loop SIMD construct will first distribute the iterations of the associated loop(s) across the implicit tasks of the parallel region in a manner consistent with any clauses that apply to the loop construct. The resulting chunks of iterations will then be converted to a SIMD loop in a manner consistent with any clauses that apply to the simd construct. The effect of any clause that applies to both constructs is as if it were applied to both constructs separately. 76 OpenMP API • Version 4.0 - July 2013 1 Restrictions 2 3 All restrictions to the loop construct and the simd construct apply to the loop SIMD construct. In addition, the following restriction applies: 4 • No ordered clause can be specified. 5 Cross References 6 • loop construct, see Section 2.7.1 on page 53. 7 • simd construct, see Section 2.8.1 on page 68. 8 • Data attribute clauses, see Section 2.14.3 on page 155. 9 10 2.9 Device Constructs 2.9.1 target data Construct 11 Summary 12 Create a device data environment for the extent of the region. 13 Syntax 14 The syntax of the target data construct is as follows: C/C++ #pragma omp target data [clause[[,] clause],...] new-line structured-block 15 where clause is one of the following: 16 device( integer-expression ) 17 map( [map-type : ] list ) 18 if( scalar-expression ) C/C++ Chapter 2 Directives 77 Fortran The syntax of the target data construct is as follows: 1 !$omp target data [clause[[,] clause],...] structured-block !$omp end target data where clause is one of the following: 2 3 device( scalar-integer-expression ) 4 map( [map-type : ] list ) 5 if( scalar-logical-expression ) The end target data directive denotes the end of the target data construct. 6 Fortran 7 Binding 8 9 The binding task region for a target data construct is the encountering task. The target region binds to the enclosing parallel or task region. 10 Description 11 12 13 14 15 16 17 When a target data construct is encountered, a new device data environment is created, and the encountering task executes the target data region. If there is no device clause, the default device is determined by the default-device-var ICV. The new device data environment is constructed from the enclosing device data environment, the data environment of the encountering task and any data-mapping clauses on the construct. When an if clause is present and the if clause expression evaluates to false, the device is the host. 18 Restrictions 19 20 • A program must not depend on any ordering of the evaluations of the clauses of the 21 22 • At most one device clause can appear on the directive. The device expression 23 • At most one if clause can appear on the directive. target data directive, or on any side effects of the evaluations of the clauses. must evaluate to a non-negative integer value. 78 OpenMP API • Version 4.0 - July 2013 1 Cross References 2 • map clause, see Section 2.14.5 on page 177. 3 • default-device-var, see Section 2.3 on page 34. 4 2.9.2 target Construct 5 Summary 6 Create a device data environment and execute the construct on the same device. 7 Syntax 8 The syntax of the target construct is as follows: C/C++ #pragma omp target [clause[[,] clause],...] new-line structured-block 9 where clause is one of the following: 10 device( integer-expression ) 11 map( [map-type : ] list ) 12 if( scalar-expression ) C/C++ Fortran 13 The syntax of the target construct is as follows: !$omp target [clause[[,] clause],...] structured-block !$omp end target 14 where clause is one of the following: 15 device( scalar-integer-expression ) 16 map( [map-type : ] list ) 17 if( scalar-logical-expression ) Chapter 2 Directives 79 The end target directive denotes the end of the target construct. 1 Fortran 2 Binding 3 4 The binding task for a target construct is the encountering task. The target region binds to the enclosing parallel or task region. 5 Description 6 7 8 9 10 11 The target construct provides a superset of the functionality and restrictions provided by the target data directive. The functionality added to the target directive is the inclusion of an executable region to be executed by a device. That is, the target directive is an executable directive. The encountering task waits for the device to complete the target region. When an if clause is present and the if clause expression evaluates to false, the target region is executed by the host device. 12 Restrictions 13 14 • If a target, target update, or target data construct appears within a target 15 16 • The result of an omp_set_default_device, omp_get_default_device, 17 18 • The effect of an access to a threadprivate variable in a target region is 19 20 21 • A variable referenced in a target construct that is not declared in the construct 22 23 • A variable referenced in a target region but not the target construct that is not region then the behavior is unspecified. or omp_get_num_devices routine called within a target region is unspecified. unspecified. is implicitly treated as if it had appeared in a map clause with a map-type of tofrom. declared in the target region must appear in a declare target directive. C++ • A throw executed inside a target region must cause execution to resume within the 24 25 same target region, and the same thread that threw the exception must catch it. C++ 26 Cross References 27 • target data construct, see Section 2.9.1 on page 77. 80 OpenMP API • Version 4.0 - July 2013 1 • default-device-var, see Section 2.3 on page 34. 2 • map clause, see Section 2.14.5 on page 177. 3 2.9.3 target update Construct 4 Summary 5 6 7 The target update directive makes the corresponding list items in the device data environment consistent with their original list items, according to the specified motion clauses. The target update construct is a stand-alone directive. 8 Syntax 9 The syntax of the target update construct is as follows: C/C++ #pragma omp target update clause[[,] clause],...] new-line 10 where motion-clause is one of the following: 11 to( list ) 12 from( list ) 13 and where clause is motion-clause or one of the following: 14 device( integer-expression ) 15 if( scalar-expression ) C/C++ Fortran 16 The syntax of the target update construct is as follows: !$omp target update clause[[,] clause],...] 17 where motion-clause is one of the following: 18 to( list ) 19 from( list ) Chapter 2 Directives 81 and where clause is motion-clause or one of the following: 1 2 device( scalar-integer-expression ) 3 if( scalar-logical-expression ) Fortran 4 Binding 5 6 The binding task for a target update construct is the encountering task.The target update directive is a stand-alone directive. 7 Description 8 9 10 11 For each list item in a to or from clause there is a corresponding list item and an original list item. If the corresponding list item is not present in the device data environment, the behavior is unspecified. Otherwise, each corresponding list item in the device data environment has an original list item in the current task's data environment. 12 13 For each list item in a from clause the value of the corresponding list item is assigned to the original list item. 14 15 For each list item in a to clause the value of the original list item is assigned to the corresponding list item. 16 The list items that appear in the to or from clauses may include array sections. 17 18 19 The device is specified in the device clause. If there is no device clause, the device is determined by the default-device-var ICV. When an if clause is present and the if clause expression evaluates to false then no assignments occur. 20 Restrictions 21 22 • A program must not depend on any ordering of the evaluations of the clauses of the 23 • At least one motion-clause must be specified. 24 • If a list item is an array section it must specify contiguous storage. 25 26 27 • A variable that is part of another variable (such as a field of a structure) but is not an 28 • A list item can only appear in a to or from clause, but not both. 29 • A list item in a to or from clause must have a mappable type. target update directive, or on any side effects of the evaluations of the clauses. array element or an array section cannot appear as a list item in a clause of a target update construct. 82 OpenMP API • Version 4.0 - July 2013 1 2 • At most one device clause can appear on the directive. The device expression 3 • At most one if clause can appear on the directive. 4 Cross References 5 • default-device-var, see Section 2.3 on page 34. 6 • target data, see Section 2.9.1 on page 77. 7 • Array sections, Section 2.4 on page 42 8 9 must evaluate to a non-negative integer value. 2.9.4 declare target Directive Summary 10 11 12 The declare target directive specifies that variables, functions (C, C++ and Fortran), and subroutines (Fortran) are mapped to a device. The declare target directive is a declarative directive. 13 Syntax 14 The syntax of the declare target directive is as follows: C/C++ #pragma omp declare target new-line declarations-definition-seq #pragma omp end declare target new-line C/C++ 15 Fortran 16 The syntax of the declare target directive is as follows: 17 For variables, functions and subroutines: !$omp declare target( list ) Chapter 2 Directives 83 1 2 where list is a comma-separated list of named variables, procedure names and named common blocks. Common block names must appear between slashes. 3 For functions and subroutines: !$omp declare target Fortran 4 5 Description 6 7 8 Variable and routine declarations that appear between the declare target and end declare target directives form an implicit list where each list item is the variable or function name. C/C++ C/C++ Fortran If a declare target does not have an explicit list, then an implicit list of one item is formed from the name of the enclosing subroutine subprogram, function subprogram or interface body to which it applies. 9 10 11 Fortran 12 13 If a list item is a function (C, C++, Fortran) or subroutine (Fortran) then a device-specific version of the routine is created that can be called from a target region. 14 15 16 17 If a list item is a variable then the original variable is mapped to a corresponding variable in the initial device data environment for all devices. If the original variable is initialized, the corresponding variable in the device data environment is initialized with the same value. 18 Restrictions 19 • A threadprivate variable cannot appear in a declare target directive. 20 • A variable declared in a declare target directive must have a mappable type. 21 22 • A variable declared in a declare target directive must be at file or namespace 23 24 • A function declared in a declare target directive must be at file, namespace, or C/C++ scope. class scope. 84 OpenMP API • Version 4.0 - July 2013 1 2 • All declarations and definitions for a function must have a declare target directive if one is specified for any of them. Otherwise, the result is unspecified. C/C++ Fortran 3 4 • If a list item is a procedure name, it must not be a generic name, procedure pointer or 5 6 • Any declare target directive with a list can only appear in a specification part 7 8 9 • Any declare target directive without a list can only appear in a specification entry name. of a subroutine subprogram, function subprogram, program or module. part of a subroutine subprogram, function subprogram or interface body to which it applies. 10 11 • If a declare target directive is specified in an interface block for a procedure, it 12 13 14 • If any procedure is declared via a procedure declaration statement, any declare 15 16 • A variable that is part of another variable (as an array or structure element) cannot 17 18 19 20 21 22 • The declare target directive must appear in the declaration section of a scoping 23 24 25 26 • If a declare target directive specifying a common block name appears in one 27 28 29 • If a declare target variable or a declare target common block is declared 30 • A blank common block cannot appear in a declare target directive. 31 32 33 • A variable can only appear in a declare target directive in the scope in which it 34 35 • A variable that appears in a declare target directive must be declared in the must match a declare target directive in the definition of the procedure. target directive with the procedure name must appear in the same specification part. appear in a declare target directive. unit in which the common block or variable is declared. Although variables in common blocks can be accessed by use association or host association, common block names cannot. This means that a common block name specified in a declare target directive must be declared to be a common block in the same scoping unit in which the declare target directive appears. program unit, then such a directive must also appear in every other program unit that contains a COMMON statement specifying the same name. It must appear after the last such COMMON statement in the program unit. with the BIND attribute, the corresponding C entities must also be specified in a declare target directive in the C program. is declared. It must not be an element of a common block or appear in an EQUIVALENCE statement. Fortran scope of a module or have the SAVE attribute, either explicitly or implicitly. Fortran Chapter 2 Directives 85 1 2.9.5 teams Construct 2 Summary 3 4 The teams construct creates a league of thread teams and the master thread of each team executes the region. 5 Syntax 6 The syntax of the teams construct is as follows: C/C++ #pragma omp teams [clause[[,] clause],...] new-line structured-block 7 where clause is one of the following: 8 num_teams( integer-expression ) 9 thread_limit( integer-expression ) 10 default(shared | none) 11 private( list ) 12 firstprivate( list ) 13 shared( list ) 14 reduction( reduction-identifier : list ) C/C++ Fortran The syntax of the teams construct is as follows: 15 !$omp teams [clause[[,] clause],...] structured-block !$omp end teams where clause is one of the following: 16 17 num_teams( scalar-integer-expression ) 18 thread_limit( scalar-integer-expression ) 86 OpenMP API • Version 4.0 - July 2013 1 default(shared | firstprivate | private | none) 2 private( list ) 3 firstprivate( list ) 4 shared( list ) 5 reduction( reduction-identifier : list ) 6 The end teams directive denotes the end of the teams construct. Fortran 7 Binding 8 The binding thread set for a teams region is the encountering thread. 9 Description 10 11 When a thread encounters a teams construct, a league of thread teams is created and the master thread of each thread team executes the teams region. 12 13 The number of teams created is implementation defined, but is less than or equal to the value specified in the num_teams clause. 14 15 16 The maximum number of threads participating in the contention group that each team initiates is implementation defined, but is less than or equal to the value specified in the thread_limit clause. 17 18 Once the teams are created, the number of teams remains constant for the duration of the teams region. 19 20 21 22 Within a teams region, team numbers uniquely identify each team. Team numbers are consecutive whole numbers ranging from zero to one less than the number of teams. A thread may obtain its own team number by a call to the omp_get_team_num library routine. 23 24 The threads other than the master thread do not begin execution until the master thread encounters a parallel region. 25 26 After the teams have completed execution of the teams region, the encountering thread resumes execution of the enclosing target region. 27 There is no implicit barrier at the end of a teams construct. Chapter 2 Directives 87 1 Restrictions 2 Restrictions to the teams construct are as follows: 3 • A program that branches into or out of a teams region is non-conforming. 4 5 • A program must not depend on any ordering of the evaluations of the clauses of the 6 7 • At most one thread_limit clause can appear on the directive. The 8 9 • At most one num_teams clause can appear on the directive. The num_teams teams directive, or on any side effects of the evaluation of the clauses. thread_limit expression must evaluate to a positive integer value. expression must evaluate to a positive integer value. 10 11 12 • If specified, a teams construct must be contained within a target construct. That 13 14 15 • distribute, parallel, parallel sections, parallel workshare, 16 Cross References: 17 • num_teams_var, see Section 2.3.5 on page 40. 18 19 • default, shared, private, firstprivate, and reduction clauses, see 20 • omp_get_num_teams routine, see Section 3.2.26 on page 221. 21 • omp_get_team_num routine, see Section 3.2.27 on page 222. 22 target construct must contain no statements or directives outside of the teams construct. and the parallel loop and parallel loop SIMD constructs are the only OpenMP constructs that can be closely nested in the teams region. Section 2.14.3 on page 155. 2.9.6 distribute Construct 23 Summary 24 25 26 27 The distribute construct specifies that the iterations of one or more loops will be executed by the thread teams in the context of their implicit tasks. The iterations are distributed across the master threads of all teams that execute the teams region to which the distribute region binds. 88 OpenMP API • Version 4.0 - July 2013 1 Syntax 2 The syntax of the distribute construct is as follows: C/C++ #pragma omp distribute [clause[[,] clause],...] new-line for-loops 3 4 Where clause is one of the following: 5 private( list ) 6 firstprivate( list ) 7 collapse( n ) 8 dist_schedule( kind[, chunk_size] ) 9 10 All associated for-loops must have the canonical form described in Section 2.6 on page 51 C/C++ Fortran 11 The syntax of the distribute construct is as follows: !$omp distribute [clause[[,] clause],...] do-loops [ !$omp end distribute ] 12 Where clause is one of the following: 13 private( list ) 14 firstprivate( list ) 15 collapse( n ) 16 dist_schedule( kind[, chunk_size] ) 17 18 If an end distribute directive is not specified, an end distribute directive is assumed at the end of the do-loop. Chapter 2 Directives 89 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. 1 2 3 4 Fortran 5 Binding 6 7 8 9 The binding thread set for a distribute region is the set of master threads created by a teams construct. A distribute region binds to the innermost enclosing teams region. Only the threads executing the binding teams region participate in the execution of the loop iterations. 10 Description 11 12 The distribute construct is associated with a loop nest consisting of one or more loops that follow the directive. 13 There is no implicit barrier at the end of a distribute construct. 14 15 16 17 18 The collapse clause may be used to specify how many loops are associated with the distribute 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 distribute construct is the one that immediately follows the distribute construct. 19 20 21 22 If more than one loop is associated with the distribute construct, then the iteration of all associated loops are collapsed into one larger iteration space. The sequential execution of the iterations in all associated loops determines the order of the iterations in the collapsed iteration space. 23 24 25 26 27 28 If dist_schedule is specified, kind must be static. If specified, iterations are divided into chunks of size chunk_size, chunks are assigned to the teams of the league in a round-robin fashion in the order of the team number. When no chunk_size is specified, the iteration space is divided into chunks that are approximately equal in size, and at most one chunk is distributed to each team of the league. Note that the size of the chunks is unspecified in this case. 29 When no dist_schedule clause is specified, the schedule is implementation defined. 30 Restrictions 31 Restrictions to the distribute construct are as follows: 90 OpenMP API • Version 4.0 - July 2013 1 • The distribute construct inherits the restrictions of the loop construct. 2 • A distribute construct must be closely nested in a teams region. 3 Cross References: 4 • loop construct, see Section 2.7.1 on page 53. 5 • teams construct, see Section 2.9.5 on page 86. 6 2.9.7 distribute simd Construct 7 Summary 8 9 The distribute simd construct specifies a loop that will be distributed across the master threads of the teams region and executed concurrently using SIMD instructions. 10 Syntax 11 The syntax of the distribute simd construct is as follows: C/C++ #pragma omp distribute simd [clause[[,] clause]...] for-loops 12 13 where clause can be any of the clauses accepted by the distribute or simd directives with identical meanings and restrictions. C/C++ 14 Fortran !$omp distribute simd [clause[[,] clause]...] do-loops [ !$omp end distribute simd ] 15 16 where clause can be any of the clauses accepted by the distribute or simd directives with identical meanings and restrictions. Chapter 2 Directives 91 If an end distribute simd directive is not specified, an end distribute simd directive is assumed at the end of the do-loops. 1 2 Fortran 3 Description 4 5 6 7 8 9 The distribute simd construct will first distribute the iterations of the associated loop(s) according to the semantics of the distribute construct and any clauses that apply to the distribute construct. The resulting chunks of iterations will then be converted to a SIMD loop in a manner consistent with any clauses that apply to the simd construct. The effect of any clause that applies to both constructs is as if it were applied to both constructs separately. 10 Restrictions 11 The restrictions for the distribute and simd constructs apply. 12 Cross References 13 • simd construct, see Section 2.8.1 on page 68. 14 • distribute construct, see Section 2.9.6 on page 88. 15 • Data attribute clauses, see Section 2.14.3 on page 155. 16 2.9.8 Distribute Parallel Loop Construct 17 Summary 18 19 The distribute parallel loop construct specifies a loop that can be executed in parallel by multiple threads that are members of multiple teams. 20 Syntax 21 The syntax of the distribute parallel loop construct is as follows: C/C++ #pragma omp distribute parallel for [clause[[,] clause]...] for-loops 92 OpenMP API • Version 4.0 - July 2013 1 2 where clause can be any of the clauses accepted by the distribute or parallel loop directives with identical meanings and restrictions. C/C++ 3 Fortran !$omp distribute parallel do [clause[[,] clause]...] do-loops [ !$omp end distribute parallel do ] 4 5 where clause can be any of the clauses accepted by the distribute or parallel loop directives with identical meanings and restrictions. 6 7 If an end distribute parallel do directive is not specified, an end distribute parallel do directive is assumed at the end of the do-loops. Fortran 8 Description 9 10 11 12 13 14 15 The distribute parallel loop construct will first distribute the iterations of the associated loop(s) according to the semantics of the distribute construct and any clauses that apply to the distribute construct. The resulting loops will then be distributed across the threads contained within the teams region to which the distribute construct binds in a manner consistent with any clauses that apply to the parallel loop construct. The effect of any clause that applies to both the distribute and parallel loop constructs is as if it were applied to both constructs separately. 16 Restrictions 17 The restrictions for the distribute and parallel loop constructs apply. 18 Cross References 19 • distribute construct, see Section 2.9.6 on page 88. 20 • Parallel loop construct, see Section 2.10.1 on page 95. 21 • Data attribute clauses, see Section Section 2.14.3 on page 155. Chapter 2 Directives 93 1 2.9.9 Distribute Parallel Loop SIMD Construct 2 Summary 3 4 5 The distribute parallel loop SIMD construct specifies a loop that can be executed concurrently using SIMD instructions in parallel by multiple threads that are members of multiple teams. 6 Syntax 7 The syntax of the distribute parallel loop SIMD construct is as follows: C/C++ #pragma omp distribute parallel for simd [clause[[,] clause]...] for-loops where clause can be any of the clauses accepted by the distribute or parallel loop SIMD directives with identical meanings and restrictions. 8 9 C/C++ Fortran The syntax of the distribute parallel loop SIMD construct is as follows: 10 !$omp distribute parallel do simd [clause[[,] clause]...] do-loops [ !$omp end distribute parallel do simd ] 11 12 where clause can be any of the clauses accepted by the distribute or parallel loop SIMD directives with identical meanings and restrictions. 13 14 If an end distribute parallel do simd directive is not specified, an end distribute parallel do simd directive is assumed at the end of the do-loops. Fortran 15 Description 16 17 18 19 The distribute parallel loop SIMD construct will first distribute the iterations of the associated loop(s) according to the semantics of the distribute construct and any clauses that apply to the distribute construct. The resulting loops will then be distributed across the threads contained within the teams region to which the 94 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 distribute construct binds in a manner consistent with any clauses that apply to the parallel loop construct. The resulting chunks of iterations will then be converted to a SIMD loop in a manner consistent with any clauses that apply to the simd construct. The effect of any clause that applies to both the distribute and parallel loop SIMD constructs is as if it were applied to both constructs separately. 6 Restrictions 7 The restrictions for the distribute and parallel loop SIMD constructs apply. 8 Cross References 9 • distribute construct, see Section 2.9.6 on page 88. 10 • Parallel loop SIMD construct, see Section 2.10.4 on page 100. 11 • Data attribute clauses, see Section Section 2.14.3 on page 155. 12 2.10 Combined Constructs 13 14 15 16 Combined constructs are shortcuts for specifying one construct immediately nested inside another construct. The semantics of the combined constructs are identical to that of explicitly specifying the first construct containing one instance of the second construct and no other statements. 17 18 19 20 21 Some combined constructs have clauses that are permitted on both constructs that were combined. Where specified, the effect is as if applying the clauses to one or both constructs. If not specified and applying the clause to one construct would result in different program behavior than applying the clause to the other construct then the program’s behavior is unspecified. 22 2.10.1 Parallel Loop Construct 23 Summary 24 25 The parallel loop construct is a shortcut for specifying a parallel construct containing one or more associated loops and no other statements. Chapter 2 Directives 95 1 Syntax 2 The syntax of the parallel loop construct is as follows: C/C++ #pragma omp parallel for [clause[[,] clause] ...] new-line for-loop where clause can be any of the clauses accepted by the parallel or for directives, except the nowait clause, with identical meanings and restrictions. 3 4 C/C++ Fortran The syntax of the parallel loop construct is as follows: 5 !$omp parallel do [clause[[,] clause] ...] do-loop [!$omp end parallel do] where clause can be any of the clauses accepted by the parallel or do directives, with identical meanings and restrictions. 6 7 If an end parallel do directive is not specified, an end parallel do directive is assumed at the end of the do-loop. nowait may not be specified on an end parallel do directive. 8 9 10 Fortran 11 Description 12 13 The semantics are identical to explicitly specifying a parallel directive immediately followed by a for directive. C/C++ C/C++ Fortran The semantics are identical to explicitly specifying a parallel directive immediately followed by a do directive, and an end do directive immediately followed by an end parallel directive. 14 15 16 Fortran 96 OpenMP API • Version 4.0 - July 2013 1 Restrictions 2 The restrictions for the parallel construct and the loop construct apply. 3 Cross References 4 • parallel construct, see Section 2.5 on page 44. 5 • loop construct, see Section 2.7.1 on page 53. 6 • Data attribute clauses, see Section 2.14.3 on page 155. 7 8 2.10.2 parallel sections Construct Summary 9 10 The parallel sections construct is a shortcut for specifying a parallel construct containing one sections construct and no other statements. 11 Syntax 12 The syntax of the parallel sections construct is as follows: C/C++ #pragma omp parallel sections [clause[[,] clause] ...] new-line { [#pragma omp section new-line] structured-block [#pragma omp section new-line structured-block ] ... } 13 14 where clause can be any of the clauses accepted by the parallel or sections directives, except the nowait clause, with identical meanings and restrictions. C/C++ Chapter 2 Directives 97 Fortran The syntax of the parallel sections construct is as follows: 1 !$omp parallel sections [clause[[,] clause] ...] [!$omp section] structured-block [!$omp section structured-block ] ... !$omp end parallel sections 2 3 where clause can be any of the clauses accepted by the parallel or sections directives, with identical meanings and restrictions. 4 5 The last section ends at the end parallel sections directive. nowait cannot be specified on an end parallel sections directive. Fortran 6 Description 7 8 The semantics are identical to explicitly specifying a parallel directive immediately followed by a sections directive. C/C++ C/C++ Fortran The semantics are identical to explicitly specifying a parallel directive immediately followed by a sections directive, and an end sections directive immediately followed by an end parallel directive. 9 10 11 Fortran 12 Restrictions 13 The restrictions for the parallel construct and the sections construct apply. 14 Cross References: 15 • parallel construct, see Section 2.5 on page 44. 16 • sections construct, see Section 2.7.2 on page 60. 17 • Data attribute clauses, see Section 2.14.3 on page 155. 98 OpenMP API • Version 4.0 - July 2013 Fortran 1 2.10.3 parallel workshare Construct 2 Summary 3 4 The parallel workshare construct is a shortcut for specifying a parallel construct containing one workshare construct and no other statements. 5 Syntax 6 The syntax of the parallel workshare construct is as follows: !$omp parallel workshare [clause[[,] clause] ...] structured-block !$omp end parallel workshare 7 8 9 where clause can be any of the clauses accepted by the parallel directive, with identical meanings and restrictions. nowait may not be specified on an end parallel workshare directive. 10 Description 11 12 13 The semantics are identical to explicitly specifying a parallel directive immediately followed by a workshare directive, and an end workshare directive immediately followed by an end parallel directive. 14 Restrictions 15 The restrictions for the parallel construct and the workshare construct apply. 16 Cross References 17 • parallel construct, see Section 2.5 on page 44. 18 • workshare construct, see Section 2.7.4 on page 65. 19 • Data attribute clauses, see Section 2.14.3 on page 155. Fortran Chapter 2 Directives 99 1 2.10.4 Parallel Loop SIMD Construct 2 Summary 3 4 The parallel loop SIMD construct is a shortcut for specifying a parallel construct containing one loop SIMD construct and no other statement. 5 Syntax C/C++ #pragma omp parallel for simd [clause[[,] clause] ...] new-line for-loops where clause can be any of the clauses accepted by the parallel, for or simd directives, except the nowait clause, with identical meanings and restrictions. 6 7 C/C++ 8 Fortran !$omp parallel do simd [clause[[,] clause] ...] do-loops !$omp end parallel do simd 9 10 where clause can be any of the clauses accepted by the parallel, do or simd directives, with identical meanings and restrictions. 11 12 13 If an end parallel do simd directive is not specified, an end parallel do simd directive is assumed at the end of the do-loop. nowait may not be specified on an end parallel do simd directive. Fortran 14 Description 15 16 17 18 The semantics of the parallel loop SIMD construct are identical to explicitly specifying a parallel directive immediately followed by a loop SIMD directive. The effect of any clause that applies to both constructs is as if it were applied to the loop SIMD construct and not to the parallel construct. 100 OpenMP API • Version 4.0 - July 2013 1 Restrictions 2 The restrictions for the parallel construct and the loop SIMD construct apply. 3 Cross References 4 • parallel construct, see Section 2.5 on page 44. 5 • loop SIMD construct, see Section 2.8.3 on page 76. 6 • Data attribute clauses, see Section 2.14.3 on page 155. 7 8 2.10.5 target teams construct Summary 9 10 The target teams construct is a shortcut for specifying a target construct containing a teams construct. 11 Syntax 12 The syntax of the target teams construct is as follows: C/C++ #pragma omp target teams [clause[[,] clause]...] structured-block 13 14 where clause can be any of the clauses accepted by the target or teams directives with identical meanings and restrictions. C/C++ 15 Fortran !$omp target teams [clause[[,] clause]...] structured-block !$omp end target teams Chapter 2 Directives 101 where clause can be any of the clauses accepted by the target or teams directives with identical meanings and restrictions. 1 2 Fortran 3 Description 4 5 The semantics are identical to explicitly specifying a target directive immediately followed by a teams directive. C/C++ C/C++ Fortran The semantics are identical to explicitly specifying a target directive immediately followed by a teams directive, and an end teams directive immediately followed by an end target directive. 6 7 8 Fortran Restrictions 9 10 The restrictions for the target and teams constructs apply. 11 Cross References 12 • target construct, see Section 2.9.2 on page 79. 13 • teams construct, see Section 2.9.5 on page 86. 14 • Data attribute clauses, see Section 2.14.3 on page 155. 15 2.10.6 teams distribute Construct 16 Summary 17 18 The teams distribute construct is a shortcut for specifying a teams construct containing a distribute construct. 102 OpenMP API • Version 4.0 - July 2013 1 Syntax 2 The syntax of the teams distribute construct is as follows: C/C++ #pragma omp teams distribute [clause[[,] clause]...] for-loops 3 4 where clause can be any of the clauses accepted by the teams or distribute directives with identical meanings and restrictions. C/C++ 5 Fortran !$omp teams distribute [clause[[,] clause]...] do-loops [ !$omp end teams distribute ] 6 7 where clause can be any of the clauses accepted by the teams or distribute directives with identical meanings and restrictions. 8 9 If an end teams distribute directive is not specified, an end teams distribute directive is assumed at the end of the do-loops. Fortran 10 Description 11 12 The semantics are identical to explicitly specifying a teams directive immediately followed by a distribute directive. Some clauses are permitted on both constructs. 13 Restrictions 14 The restrictions for the teams and distribute constructs apply. 15 Cross References 16 • teams construct, see Section 2.9.5 on page 86. 17 • distribute construct, see Section 2.9.6 on page 88. 18 • Data attribute clauses, see Section 2.14.3 on page 155. Chapter 2 Directives 103 1 2.10.7 teams distribute simd Construct 2 Summary 3 4 The teams distribute simd construct is a shortcut for specifying a teams construct containing a distribute simd construct. 5 Syntax 6 The syntax of the teams distribute simd construct is as follows: C/C++ #pragma omp teams distribute simd [clause[[,] clause]...] for-loops where clause can be any of the clauses accepted by the teams or distribute simd directives with identical meanings and restrictions. 7 8 C/C++ 9 Fortran !$omp teams distribute simd [clause[[,] clause]...] do-loops [!$omp end teams distribute simd] 10 11 where clause can be any of the clauses accepted by the teams or distribute simd directive with identical meanings and restrictions. 12 13 If an end teams distribute directive is not specified, an end teams distribute directive is assumed at the end of the do-loops. Fortran 14 Description 15 16 17 The semantics are identical to explicitly specifying a teams directive immediately followed by a distribute simd directive. Some clauses are permitted on both constructs. 104 OpenMP API • Version 4.0 - July 2013 1 Restrictions 2 The restrictions for the teams and distribute simd constructs apply. 3 Cross References 4 • teams construct, see Section 2.9.5 on page 86. 5 • distribute simd construct, see Section 2.9.7 on page 91. 6 • Data attribute clauses, see Section 2.14.3 on page 155. 7 8 2.10.8 target teams distribute Construct Summary 9 10 The target teams distribute construct is a shortcut for specifying a target construct containing a teams distribute construct. 11 Syntax 12 The syntax of the target teams distribute construct is as follows: C/C++ #pragma omp target teams distribute [clause[[,] clause]...] for-loops 13 14 where clause can be any of the clauses accepted by the target or teams distribute directives with identical meanings and restrictions. C/C++ 15 Fortran !$omp target teams distribute [clause[[,] clause]...] do-loops [ !$omp end target teams distribute ] 16 17 where clause can be any of the clauses accepted by the target or teams distribute directives with identical meanings and restrictions. Chapter 2 Directives 105 If an end target teams distribute directive is not specified, an end target teams distribute directive is assumed at the end of the do-loops. 1 2 Fortran 3 Description 4 5 The semantics are identical to explicitly specifying a target directive immediately followed by a teams distribute directive. 6 Restrictions 7 The restrictions for the target and teams distribute constructs apply. 8 Cross References 9 • target construct, see Section 2.9.1 on page 77. 10 • teams distribute construct, see Section 2.10.6 on page 102. 11 • Data attribute clauses, see Section 2.14.3 on page 155. 12 2.10.9 target teams distribute simd Construct 13 Summary 14 15 The target teams distribute simd construct is a shortcut for specifying a target construct containing a teams distribute simd construct. 16 Syntax 17 The syntax of the target teams distribute simd construct is as follows: C/C++ #pragma omp target teams distribute simd [clause[[,] clause]...] for-loops where clause can be any of the clauses accepted by the target or teams distribute simd directives with identical meanings and restrictions. 18 19 C/C++ 106 OpenMP API • Version 4.0 - July 2013 1 Fortran !$omp target teams distribute simd [clause[[,] clause]...] do-loops [!$omp end target teams distribute simd] 2 3 where clause can be any of the clauses accepted by the target or teams distribute simd directives with identical meanings and restrictions. 4 5 If an end target teams distribute simd directive is not specified, an end target teams distribute simd directive is assumed at the end of the do-loops. Fortran 6 Description 7 8 The semantics are identical to explicitly specifying a target directive immediately followed by a teams distribute simd directive. 9 Restrictions 10 The restrictions for the target and teams distribute simd constructs apply. 11 Cross References 12 • target construct, see Section 2.9.1 on page 77 13 • teams distribute simd construct, see Section 2.10.7 on page 104. 14 • Data attribute clauses, see Section 2.14.3 on page 155. 15 2.10.10 Teams Distribute Parallel Loop Construct 16 Summary 17 18 The teams distribute parallel loop construct is a shortcut for specifying a teams construct containing a distribute parallel loop construct. Chapter 2 Directives 107 1 Syntax 2 The syntax of the teams distribute parallel loop construct is as follows: C/C++ #pragma omp teams distribute parallel for [clause[[,] clause]...] for-loops where clause can be any of the clauses accepted by the teams or distribute parallel for directives with identical meanings and restrictions. 3 4 C/C++ 5 Fortran !$omp teams distribute parallel do [clause[[,] clause]...] do-loops [ !$omp end teams distribute parallel do ] 6 7 where clause can be any of the clauses accepted by the teams or distribute parallel do directives with identical meanings and restrictions. 8 9 If an end teams distribute parallel do directive is not specified, an end teams distribute parallel do directive is assumed at the end of the do-loops. Fortran 10 Description 11 12 13 The semantics are identical to explicitly specifying a teams directive immediately followed by a distribute parallel loop directive. The effect of any clause that applies to both constructs is as if it were applied to both constructs separately. 14 Restrictions 15 The restrictions for the teams and distribute parallel loop constructs apply. 16 Cross References 17 • teams construct, see Section 2.9.5 on page 86. 18 • Distribute parallel loop construct, see Section 2.9.8 on page 92. 19 • Data attribute clauses, see Section 2.14.3 on page 155. 108 OpenMP API • Version 4.0 - July 2013 2 Target Teams Distribute Parallel Loop Construct 3 Summary 4 5 The target teams distribute parallel loop construct is a shortcut for specifying a target construct containing a teams distribute parallel loop construct. 6 Syntax 7 The syntax of the target teams distribute parallel loop construct is as follows: 1 2.10.11 C/C++ #pragma omp target teams distribute parallel for [clause[[,] clause]...] for-loops 8 9 where clause can be any of the clauses accepted by the target or teams distribute parallel for directives with identical meanings and restrictions. C/C++ 10 Fortran !$omp target teams distribute parallel do [clause[[,] clause]...] do-loops [ !$omp end target teams distribute parallel do ] 11 12 where clause can be any of the clauses accepted by the target or teams distribute parallel do directives with identical meanings and restrictions. 13 14 15 If an end target teams distribute parallel do directive is not specified, an end target teams distribute parallel do directive is assumed at the end of the do-loops. Fortran 16 Description 17 18 The semantics are identical to explicitly specifying a target directive immediately followed by a teams distribute parallel loop directive. Chapter 2 Directives 109 1 Restrictions 2 The restrictions for the target and teams distribute parallel loop constructs apply. 3 Cross References 4 • target construct, see Section 2.9.2 on page 79. 5 • Distribute parallel loop construct, see Section 2.10.10 on page 107. 6 • Data attribute clauses, see Section 2.14.3 on page 155. 8 Teams Distribute Parallel Loop SIMD Construct 9 Summary 7 2.10.12 10 11 The teams distribute parallel loop SIMD construct is a shortcut for specifying a teams construct containing a distribute parallel loop SIMD construct. 12 Syntax 13 The syntax of the teams distribute parallel loop SIMD construct is as follows: C/C++ #pragma omp teams distribute parallel for simd [clause[[,] clause]...] for-loops where clause can be any of the clauses accepted by the teams or distribute parallel for simd directives with identical meanings and restrictions. 14 15 C/C++ 16 Fortran !$omp teams distribute parallel do simd [clause[[,] clause]...] do-loops [ !$omp end teams distribute parallel do simd ] 110 OpenMP API • Version 4.0 - July 2013 1 2 where clause can be any of the clauses accepted by the teams or distribute parallel do simd directives with identical meanings and restrictions. 3 4 5 If an end teams distribute parallel do simd directive is not specified, an end teams distribute parallel do simd directive is assumed at the end of the do-loops. Fortran 6 Description 7 8 9 The semantics are identical to explicitly specifying a teams directive immediately followed by a distribute parallel loop SIMD directive. The effect of any clause that applies to both constructs is as if it were applied to both constructs separately. 10 Restrictions 11 The restrictions for the teams and distribute parallel loop SIMD constructs apply. 12 Cross References 13 • teams construct, see Section 2.9.5 on page 86. 14 • Distribute parallel loop SIMD construct, see Section 2.9.9 on page 94. 15 • Data attribute clauses, see Section 2.14.3 on page 155. 17 Target Teams Distribute Parallel Loop SIMD Construct 18 Summary 19 20 The target teams distribute parallel loop SIMD construct is a shortcut for specifying a target construct containing a teams distribute parallel loop SIMD construct. 16 2.10.13 Chapter 2 Directives 111 1 Syntax 2 The syntax of the target teams distribute parallel loop SIMD construct is as follows: C/C++ #pragma omp target teams distribute parallel for simd [clause[[,] clause]...] for-loops where clause can be any of the clauses accepted by the target or teams distribute parallel for simd directives with identical meanings and restrictions. 3 4 5 C/C++ 6 Fortran !$omp target teams distribute parallel do simd [clause[[,] clause]...] do-loops [ !$omp end target teams distribute parallel do simd ] where clause can be any of the clauses accepted by the target or teams distribute parallel do simd directives with identical meanings and restrictions. 7 8 9 If an end target teams distribute parallel do simd directive is not specified, an end target teams distribute parallel do simd directive is assumed at the end of the do-loops. 10 11 12 Fortran 13 Description 14 15 The semantics are identical to explicitly specifying a target directive immediately followed by a teams distribute parallel loop SIMD directive. 16 Restrictions 17 18 The restrictions for the target and teams distribute parallel loop SIMD constructs apply. 112 OpenMP API • Version 4.0 - July 2013 1 Cross References 2 • target construct, see Section 2.9.2 on page 79. 3 • Teams distribute parallel loop SIMD construct, see Section 2.10.12 on page 110. 4 • Data attribute clauses, see Section 2.14.3 on page 155. 5 2.11 Tasking Constructs 6 2.11.1 task Construct 7 Summary 8 The task construct defines an explicit task. 9 Syntax C/C++ 10 The syntax of the task construct is as follows: #pragma omp task [clause[[,] clause] ...] new-line structured-block 11 where clause is one of the following: if(scalar-expression) final(scalar-expression) untied default(shared | none) mergeable private(list) firstprivate(list) shared(list) depend(dependence-type : list) Chapter 2 Directives 113 C/C++ 1 Fortran The syntax of the task construct is as follows: 2 !$omp task [clause[[,] clause] ...] structured-block !$omp end task where clause is one of the following: 3 if(scalar-logical-expression) final(scalar-logical-expression) untied default(private | firstprivate | shared | none) mergeable private(list) firstprivate(list) shared(list) depend(dependence-type : list) Fortran 4 5 Binding 6 7 The binding thread set of the task region is the current team. A task region binds to the innermost enclosing parallel region. 8 Description When a thread encounters a task construct, a task is generated from the code for the associated structured block. The data environment of the task is created according to the data-sharing attribute clauses on the task construct, per-data environment ICVs, and any defaults that apply. 9 10 11 12 114 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 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. 6 7 8 9 10 11 When an if clause is present on a task construct, and the if clause expression evaluates to false, an undeferred task is generated, and the encountering thread must suspend the current task region, for which execution cannot be resumed until the generated task is completed. Note that the use of a variable in an if clause expression of a task construct causes an implicit reference to the variable in all enclosing constructs. 12 13 14 15 16 When a final clause is present on a task construct and the final clause expression evaluates to true, the generated task will be a final task. All task constructs encountered during execution of a final task will generate final and included tasks. Note that the use of a variable in a final clause expression of a task construct causes an implicit reference to the variable in all enclosing constructs. 17 18 The if clause expression and the final clause expression are evaluated in the context outside of the task construct, and no ordering of those evaluations is specified. 19 20 21 22 23 24 25 A thread that encounters a task scheduling point within the task region may temporarily suspend the task region. By default, a task is tied and its suspended task region can only be resumed by the thread that started its execution. If the untied clause is present on a task construct, any thread in the team can resume the task region after a suspension. The untied clause is ignored if a final clause is present on the same task construct and the final clause expression evaluates to true, or if a task is an included task. 26 27 28 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. 29 30 31 When a mergeable clause is present on a task construct, and the generated task is an undeferred task or an included task, the implementation may generate a merged task instead. 32 33 34 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. Chapter 2 Directives 115 1 Restrictions 2 Restrictions to the task construct are as follows: 3 • A program that branches into or out of a task region is non-conforming. 4 5 • A program must not depend on any ordering of the evaluations of the clauses of the 6 • At most one if clause can appear on the directive. 7 • At most one final clause can appear on the directive. task directive, or on any side effects of the evaluations of the clauses. C++ • A throw executed inside a task region must cause execution to resume within the 8 9 same task region, and the same thread that threw the exception must catch it. C++ Fortran • Unsynchronized use of Fortran I/O statements by multiple tasks on the same unit has 10 11 unspecified behavior. Fortran 12 2.11.1.1 depend Clause 13 Summary 14 15 16 The depend clause enforces additional constraints on the scheduling of tasks. These constraints establish dependences only between sibling tasks. The clause consists of a dependence-type with one or more list items. 17 Syntax 18 The syntax of the depend clause is as follows: depend( dependence-type : list ) 19 Description 20 21 Task dependences are derived from the dependence-type of a depend clause and its list items, where dependence-type is one of the following: 116 OpenMP API • Version 4.0 - July 2013 1 2 3 The in dependence-type. The generated task will be a dependent task of all previously generated sibling tasks that reference at least one of the list items in an out or inout dependence-type list. 4 5 6 The out and inout dependence-types. The generated task will be a dependent task of all previously generated sibling tasks that reference at least one of the list items in an in, out, or inout dependence-type list. 7 The list items that appear in the depend clause may include array sections. 8 9 10 11 Note – The enforced task dependence establishes a synchronization of memory accesses performed by a dependent task with respect to accesses performed by the predecessor tasks. However, it is the responsibility of the programmer to synchronize properly with respect to other concurrent accesses that occur outside of those tasks. 12 Restrictions 13 Restrictions to the depend clause are as follows: 14 15 • List items used in depend clauses of the same task or sibling tasks must indicate 16 • List items used in depend clauses cannot be zero-length array sections. 17 18 • A variable that is part of another variable (such as a field of a structure) but is not an 19 Cross References 20 • Array sections, Section 2.4 on page 42. 21 • Task scheduling constraints, Section 2.11.3 on page 118. 22 identical storage or disjoint storage. array element or an array section cannot appear in a depend clause. 2.11.2 taskyield Construct 23 Summary 24 25 The taskyield construct specifies that the current task can be suspended in favor of execution of a different task. The taskyield construct is a stand-alone directive. Chapter 2 Directives 117 1 Syntax 2 The syntax of the taskyield construct is as follows: C/C++ #pragma omp taskyield new-line 3 C/C++ Fortran The syntax of the taskyield construct is as follows: 4 !$omp taskyield 5 Fortran 6 Binding 7 8 A taskyield region binds to the current task region. The binding thread set of the taskyield region is the current team. 9 Description 10 11 The taskyield region includes an explicit task scheduling point in the current task region. 12 Cross References 13 • Task scheduling, see Section 2.11.3 on page 118. 14 2.11.3 Task Scheduling 15 16 17 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: 18 • the point immediately following the generation of an explicit task 118 OpenMP API • Version 4.0 - July 2013 1 • after the point of completion of a task region 2 • in a taskyield region 3 • in a taskwait region 4 • at the end of a taskgroup region 5 • in an implicit and explicit barrier region 6 • the point immediately following the generation of a target region 7 • at the beginning and end of a target data region 8 • in a target update region 9 10 When a thread encounters a task scheduling point it may do one of the following, subject to the Task Scheduling Constraints (below): 11 • begin execution of a tied task bound to the current team 12 • resume any suspended task region, bound to the current team, to which it is tied 13 • begin execution of an untied task bound to the current team 14 • resume any suspended untied task region bound to the current team. 15 16 If more than one of the above choices is available, it is unspecified as to which will be chosen. 17 Task Scheduling Constraints are as follows: 18 1. An included task is executed immediately after generation of the task. 19 20 21 22 2. Scheduling of new tied tasks is constrained by the set of task regions that are currently tied to the thread, and that are not suspended in a barrier region. If this set is empty, any new tied task may be scheduled. Otherwise, a new tied task may be scheduled only if it is a descendent task of every task in the set. 23 3. A dependent task shall not be scheduled until its task dependences are fulfilled. 24 25 26 4. When an explicit task is generated by a construct containing an if clause for which the expression evaluated to false, and the previous constraints are already met, the task is executed immediately after generation of the task. 27 A program relying on any other assumption about task scheduling is non-conforming. 28 29 30 31 32 Note – Task scheduling points dynamically divide task regions into parts. Each part is executed uninterrupted from start to end. Different parts of the same task region are executed in the order in which they are encountered. In the absence of task synchronization constructs, the order in which a thread executes parts of different schedulable tasks is unspecified. 33 34 A correct program must behave correctly and consistently with all conceivable scheduling sequences that are compatible with the rules above. Chapter 2 Directives 119 1 2 3 4 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. 5 6 7 8 9 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. 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 2.12 Master and Synchronization Constructs 14 OpenMP provides the following synchronization constructs: 15 • the master construct. 16 • the critical construct. 17 • the barrier construct. 18 • the taskwait construct. 19 • the taskgroup construct. 20 • the atomic construct. 21 • the flush construct. 22 • the ordered construct. 23 2.12.1 master Construct 24 Summary 25 26 The master construct specifies a structured block that is executed by the master thread of the team. 120 OpenMP API • Version 4.0 - July 2013 1 Syntax 2 The syntax of the master construct is as follows: C/C++ #pragma omp master new-line structured-block 3 C/C++ Fortran 4 The syntax of the master construct is as follows: !$omp master structured-block !$omp end master 5 Fortran 6 Binding 7 8 9 10 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. 11 Description 12 13 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. 14 Restrictions C++ 15 16 • A throw executed inside a master region must cause execution to resume within the same master region, and the same thread that threw the exception must catch it. C++ Chapter 2 Directives 121 1 2.12.2 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 The syntax of the critical construct is as follows: C/C++ #pragma omp critical [(name)] new-line structured-block 7 C/C++ Fortran The syntax of the critical construct is as follows: 8 !$omp critical [(name)] structured-block !$omp end critical [(name)] 9 Fortran 10 Binding 11 12 13 The binding thread set for a critical region is all threads in the contention group. Region execution is restricted to a single thread at a time among all threads in the contention group, without regard to the team(s) to which the threads belong. 14 Description 15 16 17 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 in the contention group is 122 OpenMP API • Version 4.0 - July 2013 1 2 3 executing a critical region with the same name. The critical construct enforces exclusive access with respect to all critical constructs with the same name in all threads in the contention group, not just those threads in the current team. 4 5 6 Identifiers used to identify a critical construct have external linkage and are in a name space that is separate from the name spaces used by labels, tags, members, and ordinary identifiers. C/C++ C/C++ 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. 7 8 Fortran Restrictions 9 C++ • A throw executed inside a critical region must cause execution to resume within 10 11 12 the same critical region, and the same thread that threw the exception must catch it. C++ Fortran 13 The following restrictions apply to the critical construct: 14 15 • If a name is specified on a critical directive, the same name must also be 16 17 • If no name appears on the critical directive, no name can appear on the end specified on the end critical directive. critical directive. Fortran 18 2.12.3 barrier Construct 19 Summary 20 21 The barrier construct specifies an explicit barrier at the point at which the construct appears. The barrier construct is a stand-alone directive. Chapter 2 Directives 123 1 Syntax 2 The syntax of the barrier construct is as follows: C/C++ #pragma omp barrier new-line 3 C/C++ Fortran The syntax of the barrier construct is as follows: 4 !$omp barrier 5 6 Fortran 7 Binding 8 9 The binding thread set for a barrier region is the current team. A barrier region binds to the innermost enclosing parallel region. 10 Description 11 12 13 All threads of the team executing the binding parallel region must execute the barrier region and complete execution of all explicit tasks bound to this parallel region before any are allowed to continue execution beyond the barrier. 14 15 The barrier region includes an implicit task scheduling point in the current task region. 16 Restrictions 17 The following restrictions apply to the barrier construct: 18 19 • Each barrier region must be encountered by all threads in a team or by none at all, 20 21 • The sequence of worksharing regions and barrier regions encountered must be the unless cancellation has been requested for the innermost enclosing parallel region. same for every thread in a team. 124 OpenMP API • Version 4.0 - July 2013 1 2.12.4 taskwait Construct 2 Summary 3 4 The taskwait construct specifies a wait on the completion of child tasks of the current task. The taskwait construct is a stand-alone directive. 5 Syntax 6 The syntax of the taskwait construct is as follows: C/C++ #pragma omp taskwait newline 7 C/C++ Fortran 8 The syntax of the taskwait construct is as follows: !$omp taskwait 9 Fortran 10 Binding 11 12 A taskwait region binds to the current task region. The binding thread set of the taskwait region is the current team. 13 Description 14 15 16 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 all child tasks that it generated before the taskwait region complete execution. Chapter 2 Directives 125 1 2.12.5 taskgroup Construct 2 Summary 3 4 The taskgroup construct specifies a wait on completion of child tasks of the current task and their descendent tasks. 5 Syntax 6 The syntax of the taskgroup construct is as follows: C/C++ #pragma omp taskgroup new-line structured-block C/C++ 7 Fortran The syntax of the taskgroup construct is as follows: 8 !$omp taskgroup structured-block !$omp end taskgroup Fortran 9 10 Binding 11 12 A taskgroup region binds to the current task region. The binding thread set of the taskgroup region is the current team. 13 Description 14 15 16 17 When a thread encounters a taskgroup construct, it starts executing the region. There is an implicit task scheduling point at the end of the taskgroup region. The current task is suspended at the task scheduling point until all child tasks that it generated in the taskgroup region and all of their descendent tasks complete execution. 126 OpenMP API • Version 4.0 - July 2013 1 Cross References 2 • 3 2.12.6 Task scheduling, see Section 2.11.3 on page 118 atomic Construct 4 Summary 5 6 7 The atomic construct ensures that a specific storage location is accessed atomically, rather than exposing it to the possibility of multiple, simultaneous reading and writing threads that may result in indeterminate values. 8 Syntax 9 The syntax of the atomic construct takes either of the following forms: C/C++ #pragma omp atomic [read | write | update | capture] [seq_cst] new-line expression-stmt 10 or: #pragma omp atomic capture [seq_cst] new-line structured-block 11 where expression-stmt is an expression statement with one of the following forms: 12 13 • If clause is read: 14 15 • If clause is write: 16 17 18 19 20 21 22 23 • If clause is update or not present: v = x; x = expr; x++; x--; ++x; --x; x binop= expr; x = x binop expr; x = expr binop x; Chapter 2 Directives 127 C/C++ (cont.) 1 2 3 4 5 6 7 8 • If clause is capture: 9 and where structured-block is a structured block with one of the following forms: v v v v v v v = = = = = = = x++; x--; ++x; --x; x binop= expr; x = x binop expr; x = expr binop x; {v = x; x binop= expr;} {x binop= expr; v = x;} {v = x; x = x binop expr;} {v = x; x = expr binop x;} {x = x binop expr; v = x;} {x = expr binop x; v = x;} {v = x; x = expr;} {v = x; x++;} {v = x; ++x;} {++x; v = x;} {x++; v = x;} {v = x; x--;} {v = x; --x;} {--x; v = x;} {x--; v = x;} 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 In the preceding expressions: 26 • x and v (as applicable) are both l-value expressions with scalar type. 27 28 • During the execution of an atomic region, multiple syntactic occurrences of x must 29 • Neither of v and expr (as applicable) may access the storage location designated by x. 30 • Neither of x and expr (as applicable) may access the storage location designated by v. 31 • expr is an expression with scalar type. 32 • binop is one of +, *, -, /, &, ^, |, <<, or >>. 33 • binop, binop=, ++, and -- are not overloaded operators. 34 35 36 • The expression x binop expr must be numerically equivalent to x binop (expr). This designate the same storage location. requirement is satisfied if the operators in expr have precedence greater than binop, or by using parentheses around expr or subexpressions of expr. 128 OpenMP API • Version 4.0 - July 2013 1 2 3 • The expression expr binop x must be numerically equivalent to (expr) binop x. This 4 5 • For forms that allow multiple occurrences of x, the number of times that x is requirement is satisfied if the operators in expr have precedence equal to or greater than binop, or by using parentheses around expr or subexpressions of expr. evaluated is unspecified. C/C++ Fortran 6 The syntax of the atomic construct takes any of the following forms: !$omp atomic read [seq_cst] capture-statement [!$omp end atomic] 7 or !$omp atomic write [seq_cst] write-statement [!$omp end atomic] 8 or !$omp atomic [update] [seq_cst] update-statement [!$omp end atomic] 9 or !$omp atomic capture [seq_cst] update-statement capture-statement !$omp end atomic 10 or !$omp atomic capture [seq_cst] capture-statement update-statement !$omp end atomic Chapter 2 Directives 129 Fortran (cont.) or 1 !$omp atomic capture [seq_cst] capture-statement write-statement !$omp end atomic where write-statement has the following form (if clause is write): 2 x = expr 3 where capture-statement has the following form (if clause is capture or read): 4 v=x 5 and where update-statement has one of the following forms (if clause is update, capture, or not present): 6 7 8 x = x operator expr 9 x = expr operator x 10 x = intrinsic_procedure_name (x, expr_list) 11 x = intrinsic_procedure_name (expr_list, x) 12 In the preceding statements: 13 • x and v (as applicable) are both scalar variables of intrinsic type. 14 • x must not be an allocatable variable. 15 16 • During the execution of an atomic region, multiple syntactic occurrences of x must 17 18 • None of v, expr and expr_list (as applicable) may access the same storage location as 19 20 • None of x, expr and expr_list (as applicable) may access the same storage location as 21 • expr is a scalar expression. 22 23 24 • expr_list is a comma-separated, non-empty list of scalar expressions. If 25 • intrinsic_procedure_name is one of MAX, MIN, IAND, IOR, or IEOR. 26 • operator is one of +, *, -, /, .AND., .OR., .EQV., or .NEQV. . designate the same storage location. x. v. intrinsic_procedure_name refers to IAND, IOR, or IEOR, exactly one expression must appear in expr_list. 130 OpenMP API • Version 4.0 - July 2013 1 2 3 • The expression x operator expr must be numerically equivalent to x operator (expr). 4 5 6 • The expression expr operator x must be mathematically equivalent to (expr) operator 7 8 • intrinsic_procedure_name must refer to the intrinsic procedure name and not to other 9 • operator must refer to the intrinsic operator and not to a user-defined operator. This requirement is satisfied if the operators in expr have precedence greater than operator, or by using parentheses around expr or subexpressions of expr. x. This requirement is satisfied if the operators in expr have precedence equal to or greater than operator, or by using parentheses around expr or subexpressions of expr. program entities. 10 • All assignments must be intrinsic assignments. 11 12 • For forms that allow multiple occurrences of x, the number of times that x is evaluated is unspecified. Fortran 13 14 15 • In all atomic construct forms, the seq_cst clause and the clause that denotes the 16 Binding 17 18 19 20 The binding thread set for an atomic region is all threads in the contention group. atomic regions enforce exclusive access with respect to other atomic regions that access the same storage location x among all threads in the contention group without regard to the teams to which the threads belong. 21 Description 22 23 The atomic construct with the read clause forces an atomic read of the location designated by x regardless of the native machine word size. 24 25 The atomic construct with the write clause forces an atomic write of the location designated by x regardless of the native machine word size. 26 27 28 29 30 31 32 The atomic construct with the update clause forces an atomic update of the location designated by x using the designated operator or intrinsic. Note that when no clause is present, the semantics are equivalent to atomic update. Only the read and write of the location designated by x are performed mutually atomically. The evaluation of expr or expr_list need not be atomic with respect to the read or write of the location designated by x. No task scheduling points are allowed between the read and the write of the location designated by x. type of the atomic construct can appear in any order. In addition, an optional comma may be used to separate the clauses. Chapter 2 Directives 131 1 2 3 4 5 6 7 8 9 10 The atomic construct with the capture clause forces an atomic update of the location designated by x using the designated operator or intrinsic while also capturing the original or final value of the location designated by x with respect to the atomic update. The original or final value of the location designated by x is written in the location designated by v depending on the form of the atomic construct structured block or statements following the usual language semantics. Only the read and write of the location designated by x are performed mutually atomically. Neither the evaluation of expr or expr_list, nor the write to the location designated by v need be atomic with respect to the read or write of the location designated by x. No task scheduling points are allowed between the read and the write of the location designated by x. 11 12 Any atomic construct with a seq_cst clause forces the atomically performed operation to include an implicit flush operation without a list. 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 Note – As with other implicit flush regions, Section 1.4.4 on page 20 reduces the ordering that must be enforced. The intent is that, when the analogous operation exists in C++11 or C11, a sequentially consistent atomic construct has the same semantics as a memory_order_seq_cst atomic operation in C++11/C11. Similarly, a non-sequentially consistent atomic construct has the same semantics as a memory_order_relaxed atomic operation in C++11/C11. Unlike non-sequentially consistent atomic constructs, sequentially consistent atomic constructs preserve the interleaving (sequentially consistent) behavior of correct, data-race-free programs. However, they are not designed to replace the flush directive as a mechanism to enforce ordering for non-sequentially consistent atomic constructs, and attempts to do so require extreme caution. For example, a sequentially consistent atomic write construct may appear to be reordered with a subsequent non-sequentially consistent atomic write construct, since such reordering would not be observable by a correct program if the second write were outside an atomic directive. 29 30 31 32 For all forms of the atomic construct, any combination of two or more of these atomic constructs enforces mutually exclusive access to the locations designated by x. To avoid race conditions, all accesses of the locations designated by x that could potentially occur in parallel must be protected with an atomic construct. 33 34 35 36 atomic regions do not guarantee exclusive access with respect to any accesses outside of atomic regions to the same storage location x even if those accesses occur during a critical or ordered region, while an OpenMP lock is owned by the executing task, or during the execution of a reduction clause. 37 38 39 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. 132 OpenMP API • Version 4.0 - July 2013 1 2 3 A compliant implementation may enforce exclusive access between atomic regions that update different storage locations. The circumstances under which this occurs are implementation defined. 4 5 6 If the storage location designated by x is not size-aligned (that is, if the byte alignment of x is not a multiple of the size of x), then the behavior of the atomic region is implementation defined. 7 Restrictions 8 The following restriction applies to the atomic construct: C/C++ 9 10 • All atomic accesses to the storage locations designated by x throughout the program are required to have a compatible type. C/C++ Fortran 11 The following restriction applies to the atomic construct: 12 13 • All atomic accesses to the storage locations designated by x throughout the program are required to have the same type and type parameters. Fortran 14 Cross References 15 • critical construct, see Section 2.12.2 on page 122. 16 • barrier construct, see Section 2.12.3 on page 123. 17 • flush construct, see Section 2.12.7 on page 134. 18 • ordered construct, see Section 2.12.8 on page 138. 19 • reduction clause, see Section 2.14.3.6 on page 167. 20 • lock routines, see Section 3.3 on page 224. Chapter 2 Directives 133 1 2.12.7 flush Construct 2 Summary 3 4 5 6 7 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 17 for more details. The flush construct is a stand-alone directive. 8 Syntax 9 The syntax of the flush construct is as follows: C/C++ #pragma omp flush [(list)] new-line 10 C/C++ Fortran The syntax of the flush construct is as follows: 11 !$omp flush [(list)] 12 Fortran 13 Binding 14 15 16 17 18 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. 134 OpenMP API • Version 4.0 - July 2013 1 Description 2 3 4 5 6 7 A flush construct without a list, executed on a given thread, operates as if the whole thread-visible data state of the program, as defined by the base language, is flushed. A flush construct with a list applies the flush operation to the items in the list, and does not return until the operation is complete for all specified list items. An implementation may implement a flush with a list by ignoring the list, and treating it the same as a flush without a list. 8 9 If a pointer is present in the list, the pointer itself is flushed, not the memory block to which the pointer refers. C/C++ C/C++ Fortran 10 11 12 13 14 15 16 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 is of type C_PTR, the variable is flushed, but the storage that corresponds to that address is not flushed. If the list item or the subobject of the list item has the ALLOCATABLE attribute and has an allocation status of currently allocated, the allocated variable is flushed; otherwise the allocation status is flushed. Fortran Chapter 2 Directives 135 Note – Use of a flush construct with a list is extremely error prone and users are strongly discouraged from attempting it. The following examples illustrate the ordering properties of the flush operation. In the following incorrect pseudocode example, the programmer intends to prevent simultaneous execution of the protected section by the two threads, but the program does not work properly because it does not enforce the proper ordering of the operations on variables a and b. Any shared data accessed in the protected section is not guaranteed to be current or consistent during or after the protected section. The atomic notation in the pseudocode in the following two examples indicates that the accesses to a and b are ATOMIC writes and captures. Otherwise both examples would contain data races and automatically result in unspecified behavior. 1 2 3 4 5 6 7 8 9 10 Incorrect example: a = b = 0 thread 1 atomic(b = 1) flush(b) flush(a) atomic(tmp = a) if (tmp == 0) then protected section end if thread 2 atomic(a = 1) flush(a) flush(b) atomic(tmp = b) if (tmp == 0) then protected section end if The problem with this example is that operations on variables a and b are not ordered with respect to each other. For instance, nothing prevents the compiler from moving the flush of b on thread 1 or the flush of a on thread 2 to a position completely after the protected section (assuming that the protected section on thread 1 does not reference b and the protected section on thread 2 does not reference a). If either re-ordering happens, both threads can simultaneously execute the protected section. 11 12 13 14 15 16 136 OpenMP API • Version 4.0 - July 2013 1 2 3 4 The following pseudocode example correctly ensures that the protected section is executed by not more than one of the two threads at any one time. Notice that execution of the protected section by neither thread is considered correct in this example. This occurs if both flushes complete prior to either thread executing its if statement. Correct example: a = b = 0 thread 1 atomic(b = 1) flush(a,b) atomic(tmp = a) if (tmp == 0) then protected section end if thread 2 atomic(a = 1) flush(a,b) atomic(tmp = b) if (tmp == 0) then protected section end if 5 6 7 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. 8 A flush region without a list is implied at the following locations: 9 • During a barrier region. 10 • At entry to and exit from parallel, critical, and ordered regions. 11 • At exit from worksharing regions unless a nowait is present. 12 13 • At entry to and exit from the atomic operation (read, write, update, or capture) 14 • During omp_set_lock and omp_unset_lock regions. 15 16 17 • During omp_test_lock, omp_set_nest_lock, omp_unset_nest_lock 18 • Immediately before and immediately after every task scheduling point. 19 A flush region with a list is implied at the following locations: 20 21 22 23 24 • At entry to and exit from the atomic operation (read, write, update, or capture) performed in a sequentially consistent atomic region. and omp_test_nest_lock regions, if the region causes the lock to be set or unset. performed in a non-sequentially consistent atomic region, where the list contains only the storage location designated as x according to the description of the syntax of the atomic construct in Section 2.12.6 on page 127. Chapter 2 Directives 137 1 Note – A flush region is not implied at the following locations: 2 • At entry to worksharing regions. 3 • At entry to or exit from a master region. 4 2.12.8 ordered Construct 5 Summary 6 7 8 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. 9 Syntax C/C++ The syntax of the ordered construct is as follows: 10 #pragma omp ordered new-line structured-block C/C++ 11 Fortran The syntax of the ordered construct is as follows: 12 !$omp ordered structured-block !$omp end ordered 13 Fortran 138 OpenMP API • Version 4.0 - July 2013 1 Binding 2 3 4 The binding thread set for an ordered region is the current team. An ordered region binds to the innermost enclosing loop region. ordered regions that bind to different loop regions execute independently of each other. 5 Description 6 7 8 9 10 11 The threads in the team executing the loop region execute ordered regions sequentially in the order of the loop iterations. When the thread executing the first iteration of the loop encounters an ordered construct, it can enter the ordered region without waiting. When a thread executing any subsequent iteration encounters an ordered region, it waits at the beginning of that ordered region until execution of all the ordered regions belonging to all previous iterations have completed. 12 Restrictions 13 Restrictions to the ordered construct are as follows: 14 15 • The loop region to which an ordered region binds must have an ordered clause 16 17 18 • During execution of an iteration of a loop or a loop nest within a loop region, a specified on the corresponding loop (or parallel loop) construct. thread must not execute more than one ordered region that binds to the same loop region. C++ 19 20 21 • A throw executed inside a ordered region must cause execution to resume within the same ordered region, and the same thread that threw the exception must catch it. C++ 22 Cross References 23 • loop construct, see Section 2.7.1 on page 53. 24 • parallel loop construct, see Section 2.10.1 on page 95. Chapter 2 Directives 139 1 2.13 Cancellation Constructs 2 2.13.1 cancel Construct 3 Summary 4 5 The cancel construct activates cancellation of the innermost enclosing region of the type specified. The cancel construct is a stand-alone directive. 6 Syntax 7 The syntax of the cancel construct is as follows: C/C++ #pragma omp cancel construct-type-clause[[,] if-clause] new-line where construct-type-clause is one of the following 8 9 parallel 10 sections 11 for 12 taskgroup and if-clause is 13 if (scalar-expression) 14 C/C++ Fortran The syntax of the cancel construct is as follows: 15 !$omp cancel construct-type-clause[[,] if-clause] new-line where construct-type-clause is one of the following 16 17 parallel 140 OpenMP API • Version 4.0 - July 2013 1 sections 2 do 3 taskgroup 4 5 and if-clause is if (scalar-logical-expression) Fortran 6 Binding 7 8 9 10 The binding thread set of the cancel region is the current team. The cancel region binds to the innermost enclosing construct of the type corresponding to the type-clause specified in the directive (that is, the innermost parallel, sections, do, or taskgroup construct). 11 Description 12 13 14 15 The cancel construct activates cancellation of the binding construct only if cancel-var is true, in which case the construct causes the encountering task to continue execution at the end of the canceled construct. If cancel-var is false, the cancel construct is ignored. 16 17 Threads check for active cancellation only at cancellation points. Cancellation points are implied at the following locations: 18 • implicit barriers 19 • barrier regions 20 • cancel regions 21 • cancellation point regions 22 23 24 25 When a thread reaches one of the above cancellation points and if cancel-var is true, the thread immediately checks for active cancellation (that is, if cancellation has been activated by a cancel construct). If cancellation is active, the encountering thread continues execution at the end of the canceled construct. 26 27 28 29 Note – If one thread activates cancellation and another thread encounters a cancellation point, the absolute order of execution between the two threads is non-deterministic. Whether the thread that encounters a cancellation point detects the activated cancellation depends on the underlying hardware and operating system. Chapter 2 Directives 141 1 2 3 4 5 6 7 8 9 When cancellation of tasks is activated through the cancel taskgroup construct, the innermost enclosing taskgroup will be canceled. The task that encountered the cancel taskgroup construct continues execution at the end of its task region, which implies completion of that task. Any task that belongs to the innermost enclosing taskgroup and has already begun execution must run to completion or until a cancellation point is reached. Upon reaching a cancellation point and if cancellation is active, the task continues execution at the end of its taskgroup region, which implies its completion. Any task that belongs to the innermost enclosing taskgroup and that has not begun execution may be discarded, which implies its completion. 10 11 12 13 14 15 When cancellation is active for a parallel, sections, for, or do region, each thread of the binding thread set resumes execution at the end of the canceled region if a cancellation point is encountered. If the canceled region is a parallel region, any tasks that have been created by a task construct and their descendent tasks are canceled according to the above taskgroup cancellation semantics. If the canceled region is a sections, for, or do region, no task cancellation occurs. 16 The usual C++ rules for object destruction are followed when cancellation is performed. C++ C++ Fortran All private objects or subobjects with ALLOCATABLE attribute that are allocated inside the canceled construct are deallocated. 17 18 Fortran 19 20 21 Note – The user is responsible for releasing locks and similar data structures that might cause a deadlock when a cancel construct is encountered and blocked threads cannot be canceled. 22 23 If the canceled construct contains a reduction or lastprivate clause, the final value of the reduction or lastprivate variable is undefined. 24 25 26 27 When an if clause is present on a cancel construct and the if expression evaluates to false, the cancel construct does not activate cancellation. The cancellation point associated with the cancel construct is always encountered regardless of the value of the if expression. 142 OpenMP API • Version 4.0 - July 2013 1 Restrictions 2 The restrictions to the cancel construct are as follows: 3 4 • The behavior for concurrent cancellation of a region and a region nested within it is 5 6 7 8 • If construct-type-clause is taskgroup, the cancel construct must be closely 9 10 • If construct-type-clause is taskgroup and the cancel construct is not nested 11 • A worksharing construct that is canceled must not have a nowait clause. 12 • A loop construct that is canceled must not have an ordered clause. 13 14 15 • A construct that may be subject to cancellation must not encounter an orphaned 16 Cross References: 17 • cancel-var, see Section 2.3.1 on page 35 18 • cancellation point construct, see Section 2.13.2 on page 143 19 • omp_get_cancellation routine, see Section 3.2.9 on page 199 20 unspecified. nested inside a task construct. Otherwise, the cancel construct must be closely nested inside an OpenMP construct that matches the type specified in construct-type-clause of the cancel construct. inside a taskgroup region, then the behavior is unspecified. cancellation point. That is, a cancellation point must only be encountered within that construct and must not be encountered elsewhere in its region. 2.13.2 cancellation point Construct 21 Summary 22 23 24 25 The cancellation point construct introduces a user-defined cancellation point at which implicit or explicit tasks check if cancellation of the innermost enclosing region of the type specified has been activated. The cancellation point construct is a stand-alone directive. Chapter 2 Directives 143 1 Syntax 2 The syntax of the cancellation point construct is as follows: C/C++ #pragma omp cancellation point construct-type-clause new-line where construct-type-clause is one of the following 3 4 parallel 5 sections 6 for 7 taskgroup 8 C/C++ Fortran The syntax of the cancellation point construct is as follows: 9 !$omp cancellation point construct-type-clause where construct-type-clause is one of the following 10 11 parallel 12 sections 13 do 14 taskgroup 15 Fortran 16 Binding 17 A cancellation point region binds to the current task region. 144 OpenMP API • Version 4.0 - July 2013 1 Description 2 3 4 5 This directive introduces a user-defined cancellation point at which an implicit or explicit task must check if cancellation of the innermost enclosing region of the type specified in the clause has been requested. This construct does not implement a synchronization between threads or tasks. 6 7 8 9 When an implicit or explicit task reaches a user-defined cancellation point and if cancel-var is true the task immediately checks whether cancellation of the region specified in the clause has been activated. If so, the encountering task continues execution at the end of the canceled construct. 10 Restrictions 11 12 13 14 15 • A cancellation point construct for which construct-type-clause is 16 17 • An OpenMP program with orphaned cancellation point constructs is 18 Cross References: 19 • cancel-var, see Section 2.3.1 on page 35. 20 • cancel construct, see Section 2.13.1 on page 140. 21 • omp_get_cancellation routine, see Section 3.2.9 on page 199. taskgroup must be closely nested inside a task construct. A cancellation point construct for which construct-type-clause is not taskgroup must be closely nested inside an OpenMP construct that matches the type specified in construct-type-clause. non-conforming. Chapter 2 Directives 145 1 2.14 Data Environment 2 3 This section presents a directive and several clauses for controlling the data environment during the execution of parallel, task, simd, and worksharing regions. 4 5 • Section 2.14.1 on page 146 describes how the data-sharing attributes of variables 6 7 • The threadprivate directive, which is provided to create threadprivate memory, referenced in parallel, task, simd, and worksharing regions are determined. is described in Section 2.14.2 on page 150. 8 9 10 • Clauses that may be specified on directives to control the data-sharing attributes of 11 12 13 • Clauses that may be specified on directives to copy data values from private or 14 15 • Clauses that may be specified on directives to map variables to devices are described 16 variables referenced in parallel, task, simd or worksharing constructs are described in Section 2.14.3 on page 155. threadprivate variables on one thread to the corresponding variables on other threads in the team are described in Section 2.14.4 on page 173. in Section 2.14.5 on page 177. 2.14.1 Data-sharing Attribute Rules 17 18 19 This section describes how the data-sharing attributes of variables referenced in parallel, task, simd, and worksharing regions are determined. The following two cases are described separately: 20 21 • Section 2.14.1.1 on page 146 describes the data-sharing attribute rules for variables 22 23 • Section 2.14.1.2 on page 149 describes the data-sharing attribute rules for variables 24 25 referenced in a construct. referenced in a region, but outside any construct. 2.14.1.1 Data-sharing Attribute Rules for Variables Referenced in a Construct 26 27 28 The data-sharing attributes of variables that are referenced in a construct can be predetermined, explicitly determined, or implicitly determined, according to the rules outlined in this section. 29 30 31 Specifying a variable on a firstprivate, lastprivate, linear, reduction, or copyprivate clause of an enclosed construct causes an implicit reference to the variable in the enclosing construct. Specifying a variable on a map clause of an enclosed 146 OpenMP API • Version 4.0 - July 2013 1 2 3 construct may cause an implicit reference to the variable in the enclosing construct. Such implicit references are also subject to the data-sharing attribute rules outlined in this section. 4 Certain variables and objects have predetermined data-sharing attributes as follows: 5 • Variables appearing in threadprivate directives are threadprivate. 6 7 • Variables with automatic storage duration that are declared in a scope inside the 8 • Objects with dynamic storage duration are shared. 9 • Static data members are shared. C/C++ construct are private. 10 11 • The loop iteration variable(s) in the associated for-loop(s) of a for or parallel 12 13 14 • The loop iteration variable in the associated for-loop of a simd construct with just 15 16 • The loop iteration variables in the associated for-loops of a simd construct with 17 18 • Variables with static storage duration that are declared in a scope inside the construct for construct is (are) private. one associated for-loop is linear with a constant-linear-step that is the increment of the associated for-loop. multiple associated for-loops are lastprivate. are shared. C/C++ Fortran 19 20 • Variables and common blocks appearing in threadprivate directives are 21 22 • The loop iteration variable(s) in the associated do-loop(s) of a do or parallel do 23 24 25 • The loop iteration variable in the associated do-loop of a simd construct with just 26 27 • The loop iteration variables in the associated do-loops of a simd construct with 28 29 • A loop iteration variable for a sequential loop in a parallel or task construct is 30 • Implied-do indices and forall indices are private. 31 32 • Cray pointees inherit the data-sharing attribute of the storage with which their Cray 33 • Assumed-size arrays are shared. threadprivate. construct is (are) private. one associated do-loop is linear with a constant-linear-step that is the increment of the associated do-loop. multiple associated do-loops are lastprivate. private in the innermost such construct that encloses the loop. pointers are associated. Chapter 2 Directives 147 • An associate name preserves the association with the selector established at the 1 2 ASSOCIATE statement. Fortran 3 4 5 6 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. 7 8 • The loop iteration variable(s) in the associated for-loop(s) of a for or parallel 9 10 11 • The loop iteration variable in the associated for-loop of a simd construct with just 12 13 • The loop iteration variables in the associated for-loops of a simd construct with 14 15 • Variables with const-qualified type having no mutable member may be listed in a C/C++ for construct may be listed in a private or lastprivate clause. one associated for-loop may be listed in a linear clause with a constant-linear-step that is the increment of the associated for-loop. multiple associated for-loops may be listed in a lastprivate clause. firstprivate clause, even if they are static data members. C/C++ Fortran 16 17 • The loop iteration variable(s) in the associated do-loop(s) of a do or parallel do 18 19 20 • The loop iteration variable in the associated do-loop of a simd construct with just 21 22 • The loop iteration variables in the associated do-loops of a simd construct with 23 24 25 • Variables used as loop iteration variables in sequential loops in a parallel or 26 • Assumed-size arrays may be listed in a shared clause. construct may be listed in a private or lastprivate clause. one associated do-loop may be listed in a linear clause with a constant-linear-step that is the increment of the associated loop. multiple associated do-loops may be listed in a lastprivate clause. task construct may be listed in data-sharing clauses on the construct itself, and on enclosed constructs, subject to other restrictions. Fortran 27 28 Additional restrictions on the variables that may appear in individual clauses are described with each clause in Section 2.14.3 on page 155. 29 30 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. 148 OpenMP API • Version 4.0 - July 2013 1 2 3 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. 4 Rules for variables with implicitly determined data-sharing attributes are as follows: 5 6 • In a parallel or task construct, the data-sharing attributes of these variables are 7 8 • In a parallel construct, if no default clause is present, these variables are determined by the default clause, if present (see Section 2.14.3.1 on page 156). shared. 9 10 • For constructs other than task, if no default clause is present, these variables 11 12 13 • In a task construct, if no default clause is present, a variable that in the inherit their data-sharing attributes from the enclosing context. enclosing context is determined to be shared by all implicit tasks bound to the current team is shared. Fortran • In an orphaned task construct, if no default clause is present, dummy arguments 14 15 are firstprivate. Fortran 16 17 • In a task construct, if no default clause is present, a variable whose data-sharing 18 19 20 Additional restrictions on the variables for which data-sharing attributes cannot be implicitly determined in a task construct are described in Section 2.14.3.4 on page 162. 21 22 attribute is not determined by the rules above is firstprivate. 2.14.1.2 Data-sharing Attribute Rules for Variables Referenced in a Region but not in a Construct 23 24 The data-sharing attributes of variables that are referenced in a region, but not in a construct, are determined as follows: 25 26 • Variables with static storage duration that are declared in called routines in the region 27 28 • Variables with const-qualified type having no mutable member, and that are 29 30 • File-scope or namespace-scope variables referenced in called routines in the region 31 • Objects with dynamic storage duration are shared. 32 • Static data members are shared unless they appear in a threadprivate directive. C/C++ are shared. declared in called routines, are shared. are shared unless they appear in a threadprivate directive. Chapter 2 Directives 149 1 2 • Formal arguments of called routines in the region that are passed by reference inherit 3 • Other variables declared in called routines in the region are private. the data-sharing attributes of the associated actual argument. C/C++ Fortran 4 5 6 • Local variables declared in called routines in the region and that have the save 7 8 9 • Variables belonging to common blocks, or declared in modules, and referenced in attribute, or that are data initialized, are shared unless they appear in a threadprivate directive. called routines in the region are shared unless they appear in a threadprivate directive. 10 11 • Dummy arguments of called routines in the region that are passed by reference inherit 12 13 • Cray pointees inherit the data-sharing attribute of the storage with which their Cray 14 15 • Implied-do indices, forall indices, and other local variables declared in called the data-sharing attributes of the associated actual argument. pointers are associated. routines in the region are private. Fortran 16 2.14.2 threadprivate Directive 17 Summary 18 19 The threadprivate directive specifies that variables are replicated, with each thread having its own copy. The threadprivate directive is a declarative directive. 20 Syntax 21 The syntax of the threadprivate directive is as follows: C/C++ #pragma omp threadprivate(list) new-line where list is a comma-separated list of file-scope, namespace-scope, or static block-scope variables that do not have incomplete types. 22 23 C/C++ 150 OpenMP API • Version 4.0 - July 2013 Fortran 1 The syntax of the threadprivate directive is as follows: !$omp threadprivate(list) 2 3 where list is a comma-separated list of named variables and named common blocks. Common block names must appear between slashes. Fortran 4 Description 5 6 7 8 9 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. 10 11 A program in which a thread references another thread’s copy of a threadprivate variable is non-conforming. 12 13 14 The content of a threadprivate variable can change across a task scheduling point if the executing thread switches to another task that modifies the variable. For more details on task scheduling, see Section 1.3 on page 14 and Section 2.11 on page 113. 15 16 In parallel regions, references by the master thread will be to the copy of the variable in the thread that encountered the parallel region. 17 18 19 During a 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. 20 21 22 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: 23 • Neither parallel region is nested inside another explicit parallel region. 24 • The number of threads used to execute both parallel regions is the same. 25 • The thread affinity policies used to execute both parallel regions are the same. 26 27 • The value of the dyn-var internal control variable in the enclosing task region is false 28 29 30 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. at entry to both parallel regions. Chapter 2 Directives 151 C/C++ 1 2 3 4 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. 5 6 7 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. C/C++ C++ 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. 8 9 10 C++ Fortran 11 12 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. 13 14 15 16 17 18 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 variable will be implementation defined. 19 20 21 22 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: 23 24 • If it has the ALLOCATABLE attribute, each copy created will have an initial 25 • If it has the POINTER attribute: allocation status of not currently allocated. 26 27 28 • 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; 29 • otherwise, each copy created will have an association status of undefined. 30 • If it does not have either the POINTER or the ALLOCATABLE attribute: 31 32 • if it is initially defined, either through explicit initialization or default initialization, each copy created is so defined; 152 OpenMP API • Version 4.0 - July 2013 1 • otherwise, each copy created is undefined. Fortran 2 Restrictions 3 The restrictions to the threadprivate directive are as follows: 4 5 • A threadprivate variable must not appear in any clause except the copyin, 6 • A program in which an untied task accesses threadprivate storage is non-conforming. 7 8 9 • A variable that is part of another variable (as an array or structure element) cannot copyprivate, schedule, num_threads, thread_limit, and if clauses. C/C++ appear in a threadprivate clause unless it is a static data member of a C++ class. 10 11 12 • A threadprivate directive for file-scope variables must appear outside any 13 14 15 • A threadprivate directive for namespace-scope variables must appear outside 16 17 18 • Each variable in the list of a threadprivate directive at file, namespace, or class 19 20 21 • A threadprivate directive for static block-scope variables must appear in the 22 23 24 • Each variable in the list of a threadprivate directive in block scope must refer to 25 26 27 • If a variable is specified in a threadprivate directive in one translation unit, it 28 • The address of a threadprivate variable is not an address constant. definition or declaration, and must lexically precede all references to any of the variables in its list. any definition or declaration other than the namespace definition itself, and must lexically precede all references to any of the variables in its list. scope must refer to a variable declaration at file, namespace, or class scope that lexically precedes the directive. scope of the variable and not in a nested scope. The directive must lexically precede all references to any of the variables in its list. a variable declaration in the same scope that lexically precedes the directive. The variable declaration must use the static storage-class specifier. must be specified in a threadprivate directive in every translation unit in which it is declared. C/C++ C++ 29 30 31 • A threadprivate directive for static class member variables must appear in the 32 • A threadprivate variable must not have an incomplete type or a reference type. 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. Chapter 2 Directives 153 • A threadprivate variable with class type must have: 1 2 3 • an accessible, unambiguous default constructor in case of default initialization without a given initializer; 4 5 • an accessible, unambiguous constructor accepting the given argument in case of direct initialization; 6 7 • an accessible, unambiguous copy constructor in case of copy initialization with an explicit initializer. C/C++ Fortran • A variable that is part of another variable (as an array or structure element) cannot 8 9 appear in a threadprivate clause. 10 11 12 13 14 15 • The threadprivate directive must appear in the declaration section of a scoping 16 17 18 19 • If a threadprivate directive specifying a common block name appears in one 20 21 22 • If a threadprivate variable or a threadprivate common block is declared 23 • A blank common block cannot appear in a threadprivate directive. 24 25 26 • A variable can only appear in a threadprivate directive in the scope in which it 27 28 • A variable that appears in a threadprivate directive must be declared in the unit in which the common block or variable is declared. Although variables in common blocks can be accessed by use association or host association, common block names cannot. This means that a common block name specified in a threadprivate directive must be declared to be a common block in the same scoping unit in which the threadprivate directive appears. program unit, then such a directive must also appear in every other program unit that contains a COMMON statement specifying the same name. It must appear after the last such COMMON statement in the program unit. with the BIND attribute, the corresponding C entities must also be specified in a threadprivate directive in the C program. is declared. It must not be an element of a common block or appear in an EQUIVALENCE statement. scope of a module or have the SAVE attribute, either explicitly or implicitly. Fortran 29 Cross References: 30 • dyn-var ICV, see Section 2.3 on page 34. 31 • number of threads used to execute a parallel region, see Section 2.5.1 on page 47. 32 • copyin clause, see Section 2.14.4.1 on page 173. 154 OpenMP API • Version 4.0 - July 2013 1 2.14.3 Data-Sharing Attribute Clauses 2 3 4 Several constructs accept clauses that allow a user to control the data-sharing attributes of variables referenced in the construct. Data-sharing attribute clauses apply only to variables for which the names are visible in the construct on which the clause appears. 5 6 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. 7 8 9 10 11 12 Most of the clauses accept a comma-separated list of list items (see Section 2.1 on page 26). 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. C++ 13 14 15 If a variable referenced in a data-sharing attribute clause has a type derived from a template, and there are no other references to that variable in the program, then any behavior related to that variable is unspecified. C++ Fortran 16 17 18 19 20 21 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. 22 23 24 25 Although variables in common blocks can be accessed by use association or host association, common block names cannot. As a result, a common block name specified in a data-sharing attribute clause must be declared to be a common block in the same scoping unit in which the data-sharing attribute clause appears. 26 27 28 29 30 31 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. 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. Fortran Chapter 2 Directives 155 1 2.14.3.1 default clause 2 Summary 3 4 5 The default clause explicitly determines the data-sharing attributes of variables that are referenced in a parallel, task or teams construct and would otherwise be implicitly determined (see Section 2.14.1.1 on page 146). 6 Syntax 7 The syntax of the default clause is as follows: C/C++ default(shared | none) C/C++ 8 9 Fortran The syntax of the default clause is as follows: 10 default(private | firstprivate | shared | none) Fortran 11 12 Description 13 14 The default(shared) clause causes all variables referenced in the construct that have implicitly determined data-sharing attributes to be shared. Fortran 15 16 The default(firstprivate) clause causes all variables in the construct that have implicitly determined data-sharing attributes to be firstprivate. 17 18 The default(private) clause causes all variables referenced in the construct that have implicitly determined data-sharing attributes to be private. Fortran 156 OpenMP API • Version 4.0 - July 2013 1 2 3 4 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. 5 Restrictions 6 The restrictions to the default clause are as follows: 7 8 • Only a single default clause may be specified on a parallel, task, or teams 9 directive. 2.14.3.2 shared clause 10 Summary 11 12 The shared clause declares one or more list items to be shared by tasks generated by a parallel, task or teams construct. 13 Syntax 14 The syntax of the shared clause is as follows: shared(list) 15 Description 16 17 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. 18 19 20 It is the programmer's responsibility to ensure, by adding proper synchronization, that storage shared by an explicit task region does not reach the end of its lifetime before the explicit task region completes its execution. Fortran 21 22 23 24 The association status of a shared pointer becomes undefined upon entry to and on exit from the parallel, task or teams 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. Chapter 2 Directives 157 1 2 3 4 5 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. 6 7 Note – Use of intervening temporary storage may occur when the following three conditions hold regarding an actual argument in a reference to a non-intrinsic procedure: 8 a. The actual argument is one of the following: 9 • A shared variable. 10 • A subobject of a shared variable. 11 • An object associated with a shared variable. 12 • An object associated with a subobject of a shared variable. 13 b. The actual argument is also one of the following: 14 • An array section. 15 • An array section with a vector subscript. 16 • An assumed-shape array. 17 • A pointer array. 18 19 c. The associated dummy argument for this actual argument is an explicit-shape array or an assumed-size array. 20 21 22 23 These conditions effectively result in references to, and definitions of, the temporary storage during the procedure reference. Any references to (or definitions of) the shared storage that is associated with the dummy argument by any other task must be synchronized with the procedure reference to avoid possible race conditions. 24 Fortran 25 Restrictions 26 The restrictions for the shared clause are as follows: 27 28 • A variable that is part of another variable (as an array or structure element) cannot C/C++ appear in a shared clause unless it is a static data member of a C++ class. C/C++ 158 OpenMP API • Version 4.0 - July 2013 Fortran • A variable that is part of another variable (as an array or structure element) cannot 1 2 appear in a shared clause. Fortran 3 2.14.3.3 private clause 4 Summary 5 6 The private clause declares one or more list items to be private to a task or to a SIMD lane. 7 Syntax 8 The syntax of the private clause is as follows: private(list) 9 Description 10 11 12 13 14 15 16 17 18 19 20 Each task that references a list item that appears in a private clause in any statement in the construct receives a new list item. Each SIMD lane used in a simd construct that references a list item that appears in a private clause in any statement in the construct receives a new list item. Language-specific attributes for new list items are derived from the corresponding original list item. Inside the construct, all references to the original list item are replaced by references to the new list item. In the rest of the region, it is unspecified whether references are to the new list item or the original list item. Therefore, if an attempt is made to reference the original item, its value after the region is also unspecified. If a SIMD construct or a task does not reference a list item that appears in a private clause, it is unspecified whether SIMD lanes or the task receive a new list item. 21 The value and/or allocation status of the original list item will change only: 22 • if accessed and modified via pointer, 23 • if possibly accessed in the region but outside of the construct, 24 • as a side effect of directives or clauses, or Chapter 2 Directives 159 Fortran • if accessed and modified via construct association. 1 Fortran 2 3 4 List items that appear in a private, firstprivate, or reduction clause in a parallel construct may also appear in a private clause in an enclosed parallel, task, or worksharing, or simd construct. 5 6 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. 7 8 9 List items that appear in a private, firstprivate, lastprivate, or reduction clause in a worksharing construct may also appear in a private clause in an enclosed parallel or task construct. 10 11 12 13 14 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 and/or once for each SIMD lane used by the construct. 15 16 The new list item is initialized, or has an undefined initial value, as if it had been locally declared without an initializer. C/C++ C/C++ C++ 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. 17 18 19 C++ Fortran 20 21 22 23 24 25 If any statement of the construct references a list item, a new list item of the same type and type parameters is allocated: once for each implicit task in the parallel construct; once for each task generated by a task construct; and once for each SIMD lane used by a simd construct. The initial value of the new list item is undefined. Within a parallel, worksharing, task, teams, or simd region, the initial status of a private pointer is undefined. 26 For a list item or the subobject of a list item with the ALLOCATABLE attribute: 27 28 • if the allocation status is "not currently allocated", the new list item or the subobject of the new list item will have an initial allocation status of "not currently allocated"; 160 OpenMP API • Version 4.0 - July 2013 1 2 3 4 • if the allocation status is "currently allocated", the new list item or the subobject of 5 6 7 8 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: the new list item will have an initial allocation status of "currently allocated". If the new list item or the subobject of the new list item is an array, its bounds will be the same as those of the original list item or the subobject of the original list item. 9 10 • The contents, allocation, and association status of B are undefined on entry to the 11 12 • Any definition of A, or of its allocation or association status, causes the contents, 13 14 • Any definition of B, or of its allocation or association status, causes the contents, 15 16 17 18 A list item that appears in a private clause may be a selector of an ASSOCIATE construct. If the construct association is established prior to a parallel region, the association between the associate name and the original list item will be retained in the region. parallel, task, simd, or teams region. allocation, and association status of B to become undefined. allocation, and association status of A to become undefined. Fortran 19 Restrictions 20 The restrictions to the private clause are as follows: 21 22 • A variable that is part of another variable (as an array or structure element) cannot 23 24 • A variable of class type (or array thereof) that appears in a private clause requires 25 26 27 • A variable that appears in a private clause must not have a const-qualified type 28 29 • A variable that appears in a private clause must not have an incomplete type or a appear in a private clause. C/C++ an accessible, unambiguous default constructor for the class type. unless it is of class type with a mutable member. This restriction does not apply to the firstprivate clause. reference type. C/C++ Fortran 30 31 • A variable that appears in a private clause must either be definable, or an allocatable variable. This restriction does not apply to the firstprivate clause. Chapter 2 Directives 161 1 2 • Variables that appear in namelist statements, in variable format expressions, and in 3 4 • Pointers with the INTENT(IN) attribute may not appear in a private clause. This expressions for statement function definitions, may not appear in a private clause. restriction does not apply to the firstprivate clause. Fortran 5 2.14.3.4 firstprivate clause 6 Summary 7 8 9 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. 10 Syntax 11 The syntax of the firstprivate clause is as follows: firstprivate(list) 12 Description 13 14 The firstprivate clause provides a superset of the functionality provided by the private clause. 15 16 17 18 19 20 A list item that appears in a firstprivate clause is subject to the private clause semantics described in Section 2.14.3.3 on page 159, except as noted. In addition, the new list item is initialized from the original list item existing before the construct. The initialization of the new list item is done once for each task that references the list item in any statement in the construct. The initialization is done prior to the execution of the construct. 21 22 23 24 25 26 27 For a firstprivate clause on a parallel, task, or teams 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. 162 OpenMP API • Version 4.0 - July 2013 1 2 3 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. 4 5 If a list item appears in both firstprivate and lastprivate clauses, the update required for lastprivate occurs after all the initializations for firstprivate. 6 7 8 For variables of non-array type, the initialization occurs by copy assignment. For an array of elements of non-array type, each element is initialized as if by assignment from an element of the original array to the corresponding element of the new array. C/C++ C/C++ C++ 9 10 11 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++ Fortran 12 13 14 15 16 17 18 If the original list item does not have the POINTER attribute, initialization of the new list items occurs as if by intrinsic assignment, unless the original list item has the allocation status of not currently allocated, in which case the new list items will have the same status. If the original list item has the POINTER attribute, the new list items receive the same association status of the original list item as if by pointer assignment. Fortran 19 Restrictions 20 The restrictions to the firstprivate clause are as follows: 21 22 • A variable that is part of another variable (as an array or structure element) cannot 23 24 25 26 • A list item that is private within a parallel region must not appear in a appear in a firstprivate clause. firstprivate clause on a worksharing construct if any of the worksharing regions arising from the worksharing construct ever bind to any of the parallel regions arising from the parallel construct. Chapter 2 Directives 163 1 2 3 4 • A list item that is private within a teams region must not appear in a 5 6 7 8 • A list item that appears in a reduction clause of a parallel construct must not firstprivate clause on a distribute construct if any of the distribute regions arising from the distribute construct ever bind to any of the teams regions arising from the teams construct. appear in a firstprivate clause on a worksharing or task construct if any of the worksharing or task regions arising from the worksharing or task construct ever bind to any of the parallel regions arising from the parallel construct. 9 10 11 12 • A list item that appears in a reduction clause of a teams construct must not 13 14 15 • A list item that appears in a reduction clause in a worksharing construct must not appear in a firstprivate clause on a distribute construct if any of the distribute regions arising from the distribute construct ever bind to any of the teams regions arising from the teams construct. appear in a firstprivate clause in a task construct encountered during execution of any of the worksharing regions arising from the worksharing construct. C++ • A variable of class type (or array thereof) that appears in a firstprivate clause 16 17 requires an accessible, unambiguous copy constructor for the class type. C++ C/C++ • A variable that appears in a firstprivate clause must not have an incomplete 18 19 type or a reference type. C/C++ Fortran • Variables that appear in namelist statements, in variable format expressions, and in 20 21 22 expressions for statement function definitions, may not appear in a firstprivate clause. Fortran 23 2.14.3.5 lastprivate clause 24 Summary 25 26 27 The lastprivate clause declares one or more list items to be private to an implicit task or to a SIMD lane, and causes the corresponding original list item to be updated after the end of the region. 164 OpenMP API • Version 4.0 - July 2013 1 Syntax 2 The syntax of the lastprivate clause is as follows: lastprivate(list) 3 Description 4 5 The lastprivate clause provides a superset of the functionality provided by the private clause. 6 7 8 9 10 11 A list item that appears in a lastprivate clause is subject to the private clause semantics described in Section 2.14.3.3 on page 159. In addition, when a lastprivate clause appears on the directive that identifies a worksharing construct or a SIMD 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. 12 13 For an array of elements of non-array type, each element is assigned to the corresponding element of the original array. C/C++ C/C++ Fortran 14 15 If the original list item does not have the POINTER attribute, its update occurs as if by intrinsic assignment. 16 17 If the original list item has the POINTER attribute, its update occurs as if by pointer assignment. Fortran 18 19 20 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. 21 22 23 24 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. 25 26 27 28 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. Chapter 2 Directives 165 1 2 If a list item appears in both firstprivate and lastprivate clauses, the update required for lastprivate occurs after all initializations for firstprivate. 3 Restrictions 4 The restrictions to the lastprivate clause are as follows: 5 6 • A variable that is part of another variable (as an array or structure element) cannot appear in a lastprivate clause. 7 8 9 10 • A list item that is private within a parallel region, or that appears in the 11 12 13 • A variable of class type (or array thereof) that appears in a lastprivate clause 14 15 16 17 • A variable of class type (or array thereof) that appears in a lastprivate clause reduction clause of a parallel construct, must not appear in a lastprivate clause on a worksharing construct if any of the corresponding worksharing regions ever binds to any of the corresponding parallel regions. C++ requires an accessible, unambiguous default constructor for the class type, unless the list item is also specified in a firstprivate clause. requires an accessible, unambiguous copy assignment operator for the class type. The order in which copy assignment operators for different variables of class type are called is unspecified. C++ C/C++ 18 19 • A variable that appears in a lastprivate clause must not have a const-qualified 20 21 • A variable that appears in a lastprivate clause must not have an incomplete type type unless it is of class type with a mutable member. or a reference type. C/C++ Fortran 22 • A variable that appears in a lastprivate clause must be definable. 23 24 25 • An original list item with the ALLOCATABLE attribute in the sequentially last 26 27 28 • Variables that appear in namelist statements, in variable format expressions, and in iteration or lexically last section must have an allocation status of allocated upon exit from that iteration or section. expressions for statement function definitions, may not appear in a lastprivate clause. Fortran 166 OpenMP API • Version 4.0 - July 2013 1 2.14.3.6 reduction clause 2 Summary 3 4 5 6 7 The reduction clause specifies a reduction-identifier and one or more list items. For each list item, a private copy is created in each implicit task or SIMD lane, and is initialized with the initializer value of the reduction-identifier. After the end of the region, the original list item is updated with the values of the private copies using the combiner associated with the reduction-identifier. 8 Syntax 9 The syntax of the reduction clause is as follows: C/C++ reduction(reduction-identifier:list) 10 where: 11 12 reduction-identifier is either an identifier or one of the following operators: +, -, *, &, |, ^, && and || C C C++ 13 14 reduction-identifier is either an id-expression or one of the following operators: +, -, *, &, |, ^, && and || C++ 15 16 17 The following table lists each reduction-identifier that is implicitly declared at every scope for arithmetic types and its semantic initializer value. The actual initializer value is that value as expressed in the data type of the reduction list item. Identifier Initializer Combiner + omp_priv = 0 omp_out += omp_in * omp_priv = 1 omp_out *= omp_in - omp_priv = 0 omp_out += omp_in & omp_priv = ~0 omp_out &= omp_in Chapter 2 Directives 167 | omp_priv = 0 omp_out |= omp_in ^ omp_priv = 0 omp_out ^= omp_in && omp_priv = 1 omp_out = omp_in && omp_out || omp_priv = 0 omp_out = omp_in || omp_out max omp_priv = Least representable number in the reduction list item type omp_out = omp_in > omp_out ? omp_in : omp_out min omp_priv = Largest representable number in the reduction list item type omp_out = omp_in < omp_out ? omp_in : omp_out where omp_in and omp_out correspond to two identifiers that refer to storage of the type of the list item. omp_out holds the final value of the combiner operation. 1 2 3 C/C++ Fortran The syntax of the reduction clause is as follows: 4 reduction(reduction-identifier:list) 5 6 7 where reduction-identifier is either a base language identifier, or a user-defined operator, or one of the following operators: +, -, *, .and., .or., .eqv., .neqv., or one of the following intrinsic procedure names: max, min, iand, ior, ieor. 8 9 10 The following table lists each reduction-identifier that is implicitly declared for numeric and logical types and its semantic initializer value. The actual initializer value is that value as expressed in the data type of the reduction list item. Identifier 168 Initializer Combiner + omp_priv = 0 omp_out = omp_in + omp_out * omp_priv = 1 omp_out = omp_in * omp_out - omp_priv = 0 omp_out = omp_in + omp_out .and. omp_priv = .true. omp_out = omp_in .and. omp_out .or. omp_priv = .false. omp_out = omp_in .or. omp_out .eqv. omp_priv = .true. omp_out = omp_in .eqv. omp_out .neqv. omp_priv = .false. omp_out = omp_in .neqv. omp_out OpenMP API • Version 4.0 - July 2013 max omp_priv = Least representable number in the reduction list item type omp_out = max(omp_in, omp_out) min omp_priv = Largest representable number in the reduction list item type omp_out = min(omp_in, omp_out) iand omp_priv = All bits on omp_out = iand(omp_in, omp_out) ior omp_priv = 0 omp_out = ior(omp_in, omp_out) ieor omp_priv = 0 omp_out = ieor(omp_in, omp_out) 1 Fortran 2 3 4 5 Any reduction-identifier that is defined with the declare reduction directive is also valid. In that case, the initializer and combiner of the reduction-identifier are specified by the initializer-clause and the combiner in the declare reduction directive. 6 Description 7 8 The reduction clause can be used to perform some forms of recurrence calculations (involving mathematically associative and commutative operators) in parallel. 9 10 11 12 13 14 15 16 17 For parallel and worksharing constructs, a private copy of each list item is created, one for each implicit task, as if the private clause had been used. For the simd construct, a private copy of each list item is created, one for each SIMD lane as if the private clause had been used. For the teams construct, a private copy of each list item is created, one for each team in the league as if the private clause had been used. The private copy is then initialized 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 combiner of the specified reduction-identifier. 18 19 20 The reduction-identifier specified in the reduction clause must match a previously declared reduction-identifier of the same name and type for each of the list items. This match is done by means of a name lookup in the base language. C++ 21 22 If the type of a list item is a reference to a type T then the type will be considered to be T for all purposes of this clause. 23 24 If the type is a derived class, then any reduction-identifier that matches its base classes are also a match, if there is no specific match for the type. Chapter 2 Directives 169 1 2 If the reduction-identifier is not an id-expression then it is implicitly converted to one by prepending the keyword operator (for example, + becomes operator+). 3 4 If the reduction-identifier is qualified then a qualified name lookup is used to find the declaration. 5 6 If the reduction-identifier is unqualified then an argument-dependent name lookup must be performed using the type of each list item. C++ 7 8 9 10 11 12 13 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. 14 15 16 17 18 19 The location in the OpenMP program at which the values are combined and the order in which the values are combined are unspecified. Therefore, when comparing sequential and parallel runs, or when comparing one parallel run to another (even if the number of threads used is the same), there is no guarantee that bit-identical results will be obtained or that side effects (such as floating-point exceptions) will be identical or take place at the same location in the OpenMP program. 20 21 22 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. 23 Restrictions 24 The restrictions to the reduction clause are as follows: 25 26 27 • A list item that appears in a reduction clause of a worksharing construct must be 28 29 • A list item that appears in a reduction clause of the innermost enclosing 30 31 • Any number of reduction clauses can be specified on the directive, but a list item 32 33 • For a reduction-identifier declared with the declare reduction construct, the shared in the parallel regions to which any of the worksharing regions arising from the worksharing construct bind. worksharing or parallel construct may not be accessed in an explicit task. can appear only once in the reduction clauses for that directive. directive must appear before its use in a reduction clause. 170 OpenMP API • Version 4.0 - July 2013 C/C++ 1 2 3 4 5 6 7 • The type of a list item that appears in a reduction clause must be valid for the 8 • Arrays may not appear in a reduction clause. 9 • A list item that appears in a reduction clause must not be const-qualified. reduction-identifier. For a max or min reduction in C, the type of the list item must be an allowed arithmetic data type: char, int, float, double, or _Bool, possibly modified with long, short, signed, or unsigned. For a max or min reduction in C++, the type of the list item must be an allowed arithmetic data type: char, wchar_t, int, float, double, or bool, possibly modified with long, short, signed, or unsigned. 10 11 • If a list item is a reference type then it must bind to the same object for all threads of 12 • The reduction-identifier for any list item must be unambiguous and accessible. the team. C/C++ Fortran 13 14 • The type of a list item that appears in a reduction clause must be valid for the 15 • A list item that appears in a reduction clause must be definable. 16 • A procedure pointer may not appear in a reduction clause. 17 18 • A pointer with the INTENT(IN) attribute may not appear in the reduction 19 • A pointer must be associated upon entry and exit to the region. 20 • A pointer must not have its association status changed within the region. 21 22 23 • An original list item with the POINTER attribute must be associated at entry to the 24 25 26 • An original list item with the ALLOCATABLE attribute must be in the allocated state 27 28 29 • If the reduction-identifier is defined in a declare reduction directive, the 30 31 • If the reduction-identifier is a user-defined operator, the same explicit interface for reduction operator or intrinsic. clause. construct containing the reduction clause. Additionally, the list item must not be deallocated, allocated, or pointer assigned within the region. at entry to the construct containing the reduction clause. Additionally, the list item must not be deallocated and/or allocated within the region. declare reduction directive must be in the same subprogram, or accessible by host or use association. that operator must be accessible as at the declare reduction directive. Chapter 2 Directives 171 • If the reduction-identifier is defined in a declare reduction directive, any 1 2 3 4 subroutine or function referenced in the initializer clause or combiner expression must be an intrinsic function, or must have an explicit interface where the same explicit interface is accessible as at the declare reduction directive. Fortran 5 2.14.3.7 linear clause 6 Summary 7 8 The linear clause declares one or more list items to be private to a SIMD lane and to have a linear relationship with respect to the iteration space of a loop. 9 Syntax The syntax of the linear clause is as follows: 10 linear( list[:linear-step] ) 11 Description 12 13 The linear clause provides a superset of the functionality provided by the private clause. 14 15 16 17 18 19 20 A list item that appears in a linear clause is subject to the private clause semantics described in Section 2.14.3.3 on page 159 except as noted. In addition, the value of the new list item on each iteration of the associated loop(s) corresponds to the value of the original list item before entering the construct plus the logical number of the iteration times linear-step. If linear-step is not specified it is assumed to be 1. The value corresponding to the sequentially last iteration of the associated loops is assigned to the original list item. 21 Restrictions 22 23 24 • The linear-step expression must be invariant during the execution of the region 25 • A list-item cannot appear in more than one linear clause. associated with the construct. Otherwise, the execution results in unspecified behavior. 172 OpenMP API • Version 4.0 - July 2013 1 2 • A list-item that appears in a linear clause cannot appear in any other data-sharing 3 • A list-item that appears in a linear clause must be of integral or pointer type. attribute clause. C/C++ C/C++ Fortran • A list-item that appears in a linear clause must be of type integer. 4 Fortran 5 2.14.4 Data Copying Clauses 6 7 8 This section describes the copyin clause (allowed on the parallel directive and combined parallel worksharing directives) and the copyprivate clause (allowed on the single directive). 9 10 11 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. 12 13 14 15 The clauses accept a comma-separated list of list items (see Section 2.1 on page 26). 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. 16 2.14.4.1 copyin clause 17 Summary 18 19 20 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. 21 Syntax 22 The syntax of the copyin clause is as follows: copyin(list) Chapter 2 Directives 173 1 Description 2 3 4 5 6 The copy is done after the team is formed and prior to the start of execution of the associated structured block. For variables of non-array type, the copy occurs by copy assignment. For an array of elements of non-array type, each element is copied as if by assignment from an element of the master thread’s array to the corresponding element of the other thread’s array. C/C++ C/C++ C++ 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. 7 8 C++ Fortran 9 10 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. 11 12 13 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: 14 15 • If the original list item has the POINTER attribute, each copy receives the same 16 17 18 19 • If the original list item does not have the POINTER attribute, each copy becomes association status of the master thread’s copy as if by pointer assignment. defined with the value of the master thread's copy as if by intrinsic assignment, unless it has the allocation status of not currently allocated, in which case each copy will have the same status. Fortran 20 Restrictions 21 The restrictions to the copyin clause are as follows: 22 • A list item that appears in a copyin clause must be threadprivate. 23 24 • A variable of class type (or array thereof) that appears in a copyin clause requires C/C++ an accessible, unambiguous copy assignment operator for the class type. C/C++ 174 OpenMP API • Version 4.0 - July 2013 Fortran 1 2 3 • A list item that appears in a copyin clause must be threadprivate. Named variables 4 5 • A common block name that appears in a copyin clause must be declared to be a appearing in a threadprivate common block may be specified: it is not necessary to specify the whole common block. common block in the same scoping unit in which the copyin clause appears. Fortran 6 7 2.14.4.2 copyprivate clause Summary 8 9 10 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. 11 12 13 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. 14 Syntax 15 The syntax of the copyprivate clause is as follows: copyprivate(list) 16 Description 17 18 19 20 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 2.7.3 on page 63), and before any of the threads in the team have left the barrier at the end of the construct. 21 22 23 24 In all other implicit tasks belonging to the parallel region, each specified list item becomes defined with the value of the corresponding list item in the implicit task whose thread executed the structured block. For variables of non-array type, the definition occurs by copy assignment. For an array of elements of non-array type, each element is C/C++ Chapter 2 Directives 175 copied by copy assignment from an element of the array in the data environment of the implicit task associated with the thread that executed the structured block to the corresponding element of the array in the data environment of the other implicit tasks. 1 2 3 C/C++ C++ 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. 4 5 C++ Fortran 6 7 8 9 If a list item does not have the POINTER attribute, then in all other implicit tasks belonging to the parallel region, the list item becomes defined as if by intrinsic assignment with the value of the corresponding list item in the implicit task associated with the thread that executed the structured block. 10 11 12 13 If the list item has the POINTER attribute, then, in all other implicit tasks belonging to the parallel region, the list item receives, as if by pointer assignment, the same association status of the corresponding list item in the implicit task associated with the thread that executed the structured block. Fortran 14 15 16 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). 17 Restrictions 18 The restrictions to the copyprivate clause are as follows: 19 20 • All list items that appear in the copyprivate clause must be either threadprivate 21 22 • A list item that appears in a copyprivate clause may not appear in a private or or private in the enclosing context. firstprivate clause on the single construct. C++ • A variable of class type (or array thereof) that appears in a copyprivate clause 23 24 requires an accessible unambiguous copy assignment operator for the class type. C++ 176 OpenMP API • Version 4.0 - July 2013 Fortran 1 • A common block that appears in a copyprivate clause must be threadprivate. 2 3 • Pointers with the INTENT(IN) attribute may not appear in the copyprivate clause. Fortran 4 2.14.5 map Clause 5 Summary 6 7 The map clause maps a variable from the current task's data environment to the device data environment associated with the construct. 8 Syntax 9 The syntax of the map clause is as follows: map( [map-type : ] list ) 10 Description 11 The list items that appear in a map clause may include array sections. 12 13 For list items that appear in a map clause, corresponding new list items are created in the device data environment associated with the construct. 14 15 16 The original and corresponding list items may share storage such that writes to either item by one task followed by a read or write of the other item by another task without intervening synchronization can result in data races. 17 18 19 20 21 If a corresponding list item of the original list item is in the enclosing device data environment, the new device data environment uses the corresponding list item from the enclosing device data environment. No additional storage is allocated in the new device data environment and neither initialization nor assignment is performed, regardless of the map-type that is specified. Chapter 2 Directives 177 1 2 3 4 5 If a corresponding list item is not in the enclosing device data environment, a new list item with language-specific attributes is derived from the original list item and created in the new device data environment. This new list item becomes the corresponding list item to the original list item in the new device data environment. Initialization and assignment are performed if specified by the map-type. 6 7 8 9 10 If a new list item is created then a new list item of the same type, with automatic storage duration, is allocated for the construct. The storage and thus lifetime of this list item lasts until the block in which it is created exits. The size and alignment of the new list item are determined by the type of the variable. This allocation occurs if the region references the list item in any statement. 11 12 13 14 15 16 17 If the type of the variable appearing in an array section is pointer, reference to array, or reference to pointer then the variable is implicitly treated as if it had appeared in a map clause with a map-type of alloc. The corresponding variable is assigned the address of the storage location of the corresponding array section in the new device data environment. If the variable appears in a to or from clause in a target update region enclosed by the new device data environment but not as part of the specification of an array section, the behavior is unspecified. C/C++ C/C++ Fortran If a new list item is created then a new list item of the same type, type parameter, and rank is allocated. 18 19 Fortran 20 The map-type determines how the new list item is initialized. 21 22 The alloc map-type declares that on entry to the region each new corresponding list item has an undefined initial value. 23 24 The to map-type declares that on entry to the region each new corresponding list item is initialized with the original list item's value. 25 26 The from map-type declares that on exit from the region the corresponding list item's value is assigned to each original list item. 27 28 29 The tofrom map-type declares that on entry to the region each new corresponding list item is initialized with the original list item's value and that on exit from the region the corresponding list item's value is assigned to each original list item. 30 If a map-type is not specified, the map-type defaults to tofrom. 178 OpenMP API • Version 4.0 - July 2013 1 Restrictions 2 • If a list item is an array section, it must specify contiguous storage. 3 4 • At most one list item can be an array item derived from a given variable in map 5 • List items of map clauses in the same construct must not share original storage. 6 7 8 • If any part of the original storage of a list item has corresponding storage in the clauses of the same construct. enclosing device data environment, all of the original storage must have corresponding storage in the enclosing device data environment. 9 10 • A variable that is part of another variable (such as a field of a structure) but is not an 11 • If variables that share storage are mapped, the behavior is unspecified. 12 • A list item must have a mappable type. 13 • threadprivate variables cannot appear in a map clause. 14 • Initialization and assignment are through bitwise copy. 15 16 17 • A variable for which the type is pointer, reference to array, or reference to pointer and 18 19 20 • A variable for which the type is pointer, reference to array, or reference to pointer 21 22 23 • An array section derived from a variable for which the type is pointer, reference to array element or an array section cannot appear in a map clause. C/C++ an array section derived from that variable must not appear as list items of map clauses of the same construct. must not appear as a list item if the enclosing device data environment already contains an array section derived from that variable. array, or reference to pointer must not appear as a list item if the enclosing device data environment already contains that variable. C/C++ Fortran 24 25 • The value of the new list item becomes that of the original list item in the map initialization and assignment. Fortran Chapter 2 Directives 179 1 2.15 declare reduction Directive 2 Summary 3 4 5 The following section describes the directive for declaring user-defined reductions. The declare reduction directive declares a reduction-identifier that can be used in a reduction clause. The declare reduction directive is a declarative directive. 6 7 Syntax C #pragma omp declare reduction( reduction-identifier : typename-list : combiner ) [initializer-clause] new-line where: 8 9 10 • reduction-identifier is either a base language identifier or one of the following 11 • typename-list is list of type names 12 • combiner is an expression 13 14 • initializer-clause is initializer ( initializer-expr ) where initializer-expr is operators: +, -, *, &, |, ^, && and || omp_priv = initializer or function-name ( argument-list ) C 15 C++ #pragma omp declare reduction( reduction-identifier : typename-list : combiner ) [initializer-clause] new-line 16 where: 17 18 • reduction-identifier is either a base language identifier or one of the following 19 • typename-list is a list of type names 20 • combiner is an expression operators: +, -, *, &, |, ^, && and || 180 OpenMP API • Version 4.0 - July 2013 1 2 • initializer-clause is initializer ( initializer-expr ) where initializer-expr is omp_priv initializer or function-name ( argument-list ) C++ 3 Fortran !$omp declare reduction( reduction-identifier : type-list : combiner ) [initializer-clause] 4 where: 5 6 7 • reduction-identifier is either a base language identifier, or a user-defined operator, or 8 • type-list is a list of type specifiers one of the following operators: +, -, *, .and., .or., .eqv., .neqv., or one of the following intrinsic procedure names: max, min, iand, ior, ieor. 9 10 • combiner is either an assignment statement or a subroutine name followed by an 11 12 • initializer-clause is initializer ( initializer-expr ), where initializer-expr is argument list omp_priv = expression or subroutine-name ( argument-list ) Fortran 13 Description 14 15 16 17 18 19 Custom reductions can be defined using the declare reduction directive; the reduction-identifier and the type identify the declare reduction directive. The reduction-identifier can later be used in a reduction clause using variables of the type or types specified in the declare reduction directive. If the directive applies to several types then it is considered as if there were multiple declare reduction directives, one for each type. Fortran 20 21 22 23 If a type with deferred or assumed length type parameter is specified in a declare reduction directive, the reduction-identifier of that directive can be used in a reduction clause with any variable of the same type, regardless of the length type parameters with which the variable is declared. Fortran Chapter 2 Directives 181 The visibility and accessibility of this declaration are the same as those of a variable declared at the same point in the program. The enclosing context of the combiner and of the initializer-expr will be that of the declare reduction directive. The combiner and the initializer-expr must be correct in the base language as if they were the body of a function defined at the same point in the program. 1 2 3 4 5 C++ The declare reduction directive can also appear at points in the program at which a static data member could be declared. In this case, the visibility and accessibility of the declaration are the same as those of a static data member declared at the same point in the program. 6 7 8 9 C++ 10 11 12 13 14 15 The combiner specifies how partial results can be combined into a single value. The combiner can use the special variable identifiers omp_in and omp_out that are of the type of the variables being reduced with this reduction-identifier. Each of them will denote one of the values to be combined before executing the combiner. It is assumed that the special omp_out identifier will refer to the storage that holds the resulting combined value after executing the combiner. 16 17 The number of times the combiner is executed, and the order of these executions, for any reduction clause is unspecified. Fortran 18 19 If the combiner is a subroutine name with an argument list, the combiner is evaluated by calling the subroutine with the specified argument list. 20 21 If the combiner is an assignment statement, the combiner is evaluated by executing the assignment statement. Fortran 22 23 24 25 26 27 As the initializer-expr value of a user-defined reduction is not known a priori the initializer-clause can be used to specify one. Then the contents of the initializer-clause will be used as the initializer for private copies of reduction list items where the omp_priv identifier will refer to the storage to be initialized. The special identifier omp_orig can also appear in the initializer-clause and it will refer to the storage of the original variable to be reduced. 28 29 The number of times that the initializer-expr is evaluated, and the order of these evaluations, is unspecified. 182 OpenMP API • Version 4.0 - July 2013 C/C++ 1 2 3 If the initializer-expr is a function name with an argument list, the initializer-expr is evaluated by calling the function with the specified argument list. Otherwise, the initializer-expr specifies how omp_priv is declared and initialized. C/C++ C 4 5 If no initializer-clause is specified, the private variables will be initialized following the rules for initialization of objects with static storage duration. C C++ 6 7 If no initializer-expr is specified, the private variables will be initialized following the rules for default-initialization. C++ Fortran 8 9 If the initializer-expr is a subroutine name with an argument list, the initializer-expr is evaluated by calling the subroutine with the specified argument list. 10 11 If the initializer-expr is an assignment statement, the initializer-expr is evaluated by executing the assignment statement. 12 If no initializer-clause is specified, the private variables will be initialized as follows: 13 • For complex, real, or integer types, the value 0 will be used. 14 • For logical types, the value .false. will be used. 15 16 • For derived types for which default initialization is specified, default initialization 17 • Otherwise, not specifying an initializer-clause results in unspecified behavior. will be used. Fortran C/C++ 18 19 20 If reduction-identifier is used in a target region then a declare target construct must be specified for any function that can be accessed through combiner and initializer-expr. C/C++ Chapter 2 Directives 183 Fortran If reduction-identifier is used in a target region then a declare target construct must be specified for any function or subroutine that can be accessed through combiner and initializer-expr. 1 2 3 Fortran 4 Restrictions 5 • Only the variables omp_in and omp_out are allowed in the combiner. 6 • Only the variables omp_priv and omp_orig are allowed in the initializer-clause. 7 8 • If the variable omp_orig is modified in the initializer-clause, the behavior is unspecified. 9 10 • If execution of the combiner or the initializer-expr results in the execution of an 11 12 • A reduction-identifier may not be re-declared in the current scope for the same type 13 • At most one initializer-clause can be specified. 14 15 16 • A type name in a declare reduction directive cannot be a function type, an OpenMP construct or an OpenMP API call, then the behavior is unspecified. or for a type that is compatible according to the base language rules. C/C++ array type, a reference type, or a type qualified with const, volatile or restrict. C/C++ C • If the initializer-expr is a function name with an argument list, then one of the 17 18 arguments must be the address of omp_priv. C C++ • If the initializer-expr is a function name with an argument list, then one of the 19 20 arguments must be omp_priv or the address of omp_priv. C++ Fortran • If the initializer-expr is a subroutine name with an argument list, then one of the 21 22 arguments must be omp_priv. 184 OpenMP API • Version 4.0 - July 2013 1 2 3 4 • If the declare reduction directive appears in a module and the corresponding 5 6 7 • If the reduction-identifier is a user-defined operator or an extended operator, the reduction clause does not appear in the same module, the reduction-identifier must be a user-defined operator, one of the allowed operators or one of the allowed intrinsic procedures. interface for that operator must be defined in the same subprogram, or must be accessible by host or use association. 8 9 10 11 12 • If the declare reduction directive appears in a module, any user-defined 13 14 15 • Any subroutine or function used in the initializer clause or combiner 16 17 • If the length type parameter is specified for a character type, it must be a constant, a 18 19 20 • If a character type with deferred or assumed length parameter is specified in a operators used in the combiner must be defined in the same subprogram, or must be accessible by host or use association. The user-defined operators must also be accessible by host or use association in the subprogram in which the corresponding reduction clause appears. expression must be an intrinsic function, or must have an explicit interface in the same subprogram or must be accessible by host or use association. colon or an *. declare reduction directive, no other declare reduction directives with character type and the same reduction-identifier are allowed in the same scope. Fortran 21 Cross References 22 • reduction clause, Section 2.14.3.6 on page 167. Chapter 2 Directives 185 1 2.16 Nesting of Regions 2 3 This section describes a set of restrictions on the nesting of regions. The restrictions on nesting are as follows: 4 5 • A worksharing region may not be closely nested inside a worksharing, explicit task, 6 7 • A barrier region may not be closely nested inside a worksharing, explicit task, 8 9 • A master region may not be closely nested inside a worksharing, atomic, or critical, ordered, atomic, or master region. critical, ordered, atomic, or master region. explicit task region. 10 11 • An ordered region may not be closely nested inside a critical, atomic, or 12 13 • An ordered region must be closely nested inside a loop region (or parallel loop 14 15 16 • A critical region may not be nested (closely or otherwise) inside a critical 17 • OpenMP constructs may not be nested inside an atomic region. 18 • OpenMP constructs may not be nested inside a simd region. 19 20 • If a target, target update, or target data construct appears within a 21 22 23 • If specified, a teams construct must be contained within a target construct. That 24 25 26 • distribute, parallel, parallel sections, parallel workshare, and 27 • A distribute construct must be closely nested in a teams region. 28 29 30 31 32 • If construct-type-clause is taskgroup, the cancel construct must be closely 33 34 35 36 • A cancellation point construct for which construct-type-clause is explicit task region. region) with an ordered clause. region with the same name. Note that this restriction is not sufficient to prevent deadlock. target region then the behavior is unspecified. target construct must contain no statements or directives outside of the teams construct. the parallel loop and parallel loop SIMD constructs are the only OpenMP constructs that can be closely nested in the teams region. nested inside a task construct and the cancel construct must be nested inside a taskgroup region. Otherwise, the cancel construct must be closely nested inside an OpenMP construct that matches the type specified in construct-type-clause of the cancel construct. taskgroup must be nested inside a task construct. A cancellation point construct for which construct-type-clause is not taskgroup must be closely nested inside an OpenMP construct that matches the type specified in construct-type-clause. 186 OpenMP API • Version 4.0 - July 2013 1 CHAPTER 3 2 Runtime Library Routines 3 4 This chapter describes the OpenMP API runtime library routines and is divided into the following sections: 5 • Runtime library definitions (Section 3.1 on page 188). 6 7 • Execution environment routines that can be used to control and to query the parallel 8 9 • Lock routines that can be used to synchronize access to data (Section 3.3 on page 10 execution environment (Section 3.2 on page 189). 224). • Portable timer routines (Section 3.4 on page 233). 11 12 13 Throughout this chapter, true and false are used as generic terms to simplify the description of the routines. 14 true means a nonzero integer value and false means an integer value of zero. C/C++ C/C++ Fortran 15 true means a logical value of .TRUE. and false means a logical value of .FALSE.. Fortran Fortran 16 Restrictions 17 The following restriction applies to all OpenMP runtime library routines: 18 19 • OpenMP runtime library routines may not be called from PURE or ELEMENTAL procedures. Fortran 187 1 3.1 Runtime Library Definitions 2 3 4 5 6 7 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, schedule, and thread affinity policy data types. In addition, each set of definitions may specify other implementation specific values. 8 The library routines are external functions with “C” linkage. C/C++ 9 10 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: 11 • The prototypes of all the routines in the chapter. 12 • The type omp_lock_t. 13 • The type omp_nest_lock_t. 14 • The type omp_sched_t. 15 • The type omp_proc_bind_t. 16 See Section C.1 on page 288 for an example of this file. C/C++ Fortran 17 18 The OpenMP Fortran API runtime library routines are external procedures. The return values of these routines are of default kind, unless otherwise specified. 19 20 21 22 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. 23 These files define the following: 24 • The interfaces of all of the routines in this chapter. 25 • The integer parameter omp_lock_kind. 26 • The integer parameter omp_nest_lock_kind. 27 • The integer parameter omp_sched_kind. 28 • The integer parameter omp_proc_bind_kind. 188 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 • The integer parameter openmp_version with a value yyyymm where yyyy 6 See Section C.2 on page 290 and Section C.3 on page 293 for examples of these files. 7 8 9 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 C.4 for an example of such an extension. 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 32). Fortran 10 3.2 This section describes routines that affect and monitor threads, processors, and the parallel environment. 11 12 13 Execution Environment Routines 3.2.1 omp_set_num_threads 14 Summary 15 16 17 The omp_set_num_threads routine affects the number of threads to be used for subsequent parallel regions that do not specify a num_threads clause, by setting the value of the first element of the nthreads-var ICV of the current task. Chapter 3 Runtime Library Routines 189 Format 1 C/C++ void omp_set_num_threads(int num_threads); C/C++ 2 Fortran subroutine omp_set_num_threads(num_threads) integer num_threads Fortran 3 4 Constraints on Arguments 5 6 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. 7 Binding 8 The binding task set for an omp_set_num_threads region is the generating task. 9 Effect 10 11 The effect of this routine is to set the value of the first element of the nthreads-var ICV of the current task to the value specified in the argument. 12 13 See Section 2.5.1 on page 47 for the rules governing the number of threads used to execute a parallel region. 14 Cross References 15 • nthreads-var ICV, see Section 2.3 on page 34. 16 • OMP_NUM_THREADS environment variable, see Section 4.2 on page 239. 17 • omp_get_max_threads routine, see Section 3.2.3 on page 192. 18 • parallel construct, see Section 2.5 on page 44. 19 • num_threads clause, see Section 2.5 on page 44. 190 OpenMP API • Version 4.0 - July 2013 1 3.2.2 omp_get_num_threads 2 Summary 3 4 The omp_get_num_threads routine returns the number of threads in the current team. 5 6 Format C/C++ int omp_get_num_threads(void); C/C++ 7 Fortran integer function omp_get_num_threads() Fortran 8 9 Binding 10 11 The binding region for an omp_get_num_threads region is the innermost enclosing parallel region. 12 Effect 13 14 15 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. 16 17 See Section 2.5.1 on page 47 for the rules governing the number of threads used to execute a parallel region. Chapter 3 Runtime Library Routines 191 1 Cross References 2 • parallel construct, see Section 2.5 on page 44. 3 • omp_set_num_threads routine, see Section 3.2.1 on page 189. 4 • OMP_NUM_THREADS environment variable, see Section 4.2 on page 239. 5 3.2.3 omp_get_max_threads 6 Summary 7 8 9 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 construct without a num_threads clause were encountered after execution returns from this routine. Format 10 C/C++ int omp_get_max_threads(void); C/C++ 11 Fortran integer function omp_get_max_threads() Fortran 12 13 Binding 14 The binding task set for an omp_get_max_threads region is the generating task. 192 OpenMP API • Version 4.0 - July 2013 1 Effect 2 3 4 5 The value returned by omp_get_max_threads is the value of the first element of the nthreads-var ICV of the current task. This value is also an upper bound on the number of threads that could be used to form a new team if a parallel region without a num_threads clause were encountered after execution returns from this routine. 6 7 See Section 2.5.1 on page 47 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 Cross References 12 • nthreads-var ICV, see Section 2.3 on page 34. 13 • parallel construct, see Section 2.5 on page 44. 14 • num_threads clause, see Section 2.5 on page 44. 15 • omp_set_num_threads routine, see Section 3.2.1 on page 189. 16 • OMP_NUM_THREADS environment variable, see Section 4.2 on page 239. 17 3.2.4 omp_get_thread_num 18 Summary 19 20 The omp_get_thread_num routine returns the thread number, within the current team, of the calling thread. Chapter 3 Runtime Library Routines 193 Format 1 C/C++ int omp_get_thread_num(void); C/C++ 2 Fortran integer function omp_get_thread_num() Fortran 3 4 Binding 5 6 7 The binding thread set for an omp_get_thread_num region is the current team. The binding region for an omp_get_thread_num region is the innermost enclosing parallel region. 8 Effect 9 10 11 12 13 The omp_get_thread_num routine returns the thread number of the calling thread, within the team executing the parallel region to which the routine region binds. The thread number is an integer between 0 and one less than the value returned by omp_get_num_threads, inclusive. The thread number of the master thread of the team is 0. The routine returns 0 if it is called from the sequential part of a program. 14 15 16 Note – The thread number may change during the execution of an untied task. The value returned by omp_get_thread_num is not generally useful during the execution of such a task region. 17 Cross References 18 • omp_get_num_threads routine, see Section 3.2.2 on page 191. 194 OpenMP API • Version 4.0 - July 2013 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 device. 5 Format C/C++ int omp_get_num_procs(void); C/C++ 6 Fortran integer function omp_get_num_procs() Fortran 7 8 Binding 9 10 11 The binding thread set for an omp_get_num_procs region is all threads on a device. The effect of executing this routine is not related to any specific region corresponding to any construct or API routine. 12 Effect 13 14 15 16 17 The omp_get_num_procs routine returns the number of processors that are available to the device at the time the routine is called. Note that this value may change between the time that it is determined by the omp_get_num_procs routine and the time that it is read in the calling context due to system actions outside the control of the OpenMP implementation. Chapter 3 Runtime Library Routines 195 1 3.2.6 omp_in_parallel 2 Summary 3 4 The omp_in_parallel routine returns true if the active-levels-var ICV is greater than zero; otherwise, it returns false. 5 Format C/C++ int omp_in_parallel(void); C/C++ 6 Fortran logical function omp_in_parallel() Fortran 7 8 Binding 9 The binding task set for an omp_in_parallel region is the generating task. 10 Effect 11 12 13 The effect of the omp_in_parallel routine is to return true if the current task is enclosed by an active parallel region, and the parallel region is enclosed by the outermost initial task region on the device; otherwise it returns false. 14 Cross References 15 • active-levels-var, see Section 2.3 on page 34. 16 • omp_get_active_level routine, see Section 3.2.20 on page 214. 196 OpenMP API • Version 4.0 - July 2013 1 3.2.7 omp_set_dynamic 2 Summary 3 4 5 The omp_set_dynamic routine enables or disables dynamic adjustment of the number of threads available for the execution of subsequent parallel regions by setting the value of the dyn-var ICV. 6 Format C/C++ void omp_set_dynamic(int dynamic_threads); C/C++ 7 Fortran subroutine omp_set_dynamic (dynamic_threads) logical dynamic_threads Fortran 8 9 Binding 10 The binding task set for an omp_set_dynamic region is the generating task. 11 Effect 12 13 14 15 16 For implementations that support dynamic adjustment of the number of threads, if the argument to omp_set_dynamic evaluates to true, dynamic adjustment is enabled for the current task; otherwise, dynamic adjustment is disabled for the current task. For implementations that do not support dynamic adjustment of the number of threads this routine has no effect: the value of dyn-var remains false. 17 18 See Section 2.5.1 on page 47 for the rules governing the number of threads used to execute a parallel region. Chapter 3 Runtime Library Routines 197 1 Cross References: 2 • dyn-var ICV, see Section 2.3 on page 34. 3 • omp_get_num_threads routine, see Section 3.2.2 on page 191. 4 • omp_get_dynamic routine, see Section 3.2.8 on page 198. 5 • OMP_DYNAMIC environment variable, see Section 4.3 on page 240. 6 3.2.8 omp_get_dynamic 7 Summary 8 9 The omp_get_dynamic routine returns the value of the dyn-var ICV, which determines whether dynamic adjustment of the number of threads is enabled or disabled. Format 10 C/C++ int omp_get_dynamic(void); C/C++ 11 Fortran logical function omp_get_dynamic() Fortran 12 13 Binding 14 The binding task set for an omp_get_dynamic region is the generating task. 15 Effect 16 17 18 This routine returns true if dynamic adjustment of the number of threads is enabled for the current task; it returns false, otherwise. If an implementation does not support dynamic adjustment of the number of threads, then this routine always returns false. 198 OpenMP API • Version 4.0 - July 2013 1 2 See Section 2.5.1 on page 47 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 34. 5 • omp_set_dynamic routine, see Section 3.2.7 on page 197. 6 • OMP_DYNAMIC environment variable, see Section 4.3 on page 240. 7 8 3.2.9 omp_get_cancellation Summary 9 10 The omp_get_cancellation routine returns the value of the cancel-var ICV, which controls the behavior of the cancel construct and cancellation points. 11 Format C/C++ 12 int omp_get_cancellation(void); 13 C/C++ Fortran 14 logical function omp_get_cancellation() 15 Fortran 16 Binding 17 The binding task set for an omp_get_cancellation region is the whole program. Chapter 3 Runtime Library Routines 199 1 Effect 2 This routine returns true if cancellation is activated. It returns false otherwise. 3 Cross References: 4 • cancel-var ICV, see Section 2.3.1 on page 35. 5 • OMP_CANCELLATION environment variable, see Section 4.11 on page 246. 6 3.2.10 omp_set_nested 7 Summary 8 9 The omp_set_nested routine enables or disables nested parallelism, by setting the nest-var ICV. Format 10 C/C++ void omp_set_nested(int nested); C/C++ 11 Fortran subroutine omp_set_nested (nested) logical nested Fortran 12 200 OpenMP API • Version 4.0 - July 2013 1 Binding 2 The binding task set for an omp_set_nested region is the generating task. 3 Effect 4 5 6 7 8 For implementations that support nested parallelism, if the argument to omp_set_nested evaluates to true, nested parallelism is enabled for the current task; otherwise, nested parallelism is disabled for the current task. For implementations that do not support nested parallelism, this routine has no effect: the value of nest-var remains false. 9 10 See Section 2.5.1 on page 47 for the rules governing the number of threads used to execute a parallel region. 11 Cross References 12 • nest-var ICV, see Section 2.3 on page 34. 13 • omp_set_max_active_levels routine, see Section 3.2.15 on page 207. 14 • omp_get_max_active_levels routine, see Section 3.2.16 on page 209. 15 • omp_get_nested routine, see Section 3.2.11 on page 201. 16 • OMP_NESTED environment variable, see Section 4.6 on page 243. 17 3.2.11 omp_get_nested 18 Summary 19 20 The omp_get_nested routine returns the value of the nest-var ICV, which determines if nested parallelism is enabled or disabled. Chapter 3 Runtime Library Routines 201 Format 1 C/C++ int omp_get_nested(void); C/C++ 2 Fortran logical function omp_get_nested() Fortran 3 4 Binding 5 The binding task set for an omp_get_nested region is the generating task. 6 Effect 7 8 9 This routine returns true if nested parallelism is enabled for the current task; it returns false, otherwise. If an implementation does not support nested parallelism, this routine always returns false. 10 11 See Section 2.5.1 on page 47 for the rules governing the number of threads used to execute a parallel region. 12 Cross References 13 • nest-var ICV, see Section 2.3 on page 34. 14 • omp_set_nested routine, see Section 3.2.10 on page 200. 15 • OMP_NESTED environment variable, see Section 4.6 on page 243. 202 OpenMP API • Version 4.0 - July 2013 1 3.2.12 omp_set_schedule 2 Summary 3 4 The omp_set_schedule routine affects the schedule that is applied when runtime is used as schedule kind, by setting the value of the run-sched-var ICV. 5 Format 6 C/C++ void omp_set_schedule(omp_sched_t kind, int modifier); 7 C/C++ 8 Fortran subroutine omp_set_schedule(kind, modifier) integer (kind=omp_sched_kind) kind integer modifier Fortran 9 10 Constraints on Arguments 11 12 13 14 15 The first argument passed to this routine can be one of the valid OpenMP schedule kinds (except for runtime) or any implementation specific schedule. The C/C++ header file (omp.h) and the Fortran include file (omp_lib.h) and/or Fortran 90 module file (omp_lib) define the valid constants. The valid constants must include the following, which can be extended with implementation specific values: Chapter 3 Runtime Library Routines 203 C/C++ 1 typedef enum omp_sched_t { omp_sched_static = 1, omp_sched_dynamic = 2, omp_sched_guided = 3, omp_sched_auto = 4 } omp_sched_t; C/C++ 2 Fortran integer(kind=omp_sched_kind), integer(kind=omp_sched_kind), integer(kind=omp_sched_kind), integer(kind=omp_sched_kind), parameter parameter parameter parameter :: :: :: :: omp_sched_static = 1 omp_sched_dynamic = 2 omp_sched_guided = 3 omp_sched_auto = 4 Fortran 3 4 Binding 5 The binding task set for an omp_set_schedule region is the generating task. 6 Effect The effect of this routine is to set the value of the run-sched-var ICV of the current task to the values specified in the two arguments. The schedule is set to the schedule type specified by the first argument kind. It can be any of the standard schedule types or any other implementation specific one. For the schedule types static, dynamic, and guided the chunk_size is set to the value of the second argument, or to the default chunk_size if the value of the second argument is less than 1; for the schedule type auto the second argument has no meaning; for implementation specific schedule types, the values and associated meanings of the second argument are implementation defined. 7 8 9 10 11 12 13 14 204 OpenMP API • Version 4.0 - July 2013 1 Cross References 2 • run-sched-var ICV, see Section 2.3 on page 34. 3 • omp_get_schedule routine, see Section 3.2.13 on page 205. 4 • OMP_SCHEDULE environment variable, see Section 4.1 on page 238. 5 • Determining the schedule of a worksharing loop, see Section 2.7.1.1 on page 59. 6 3.2.13 omp_get_schedule 7 Summary 8 9 The omp_get_schedule routine returns the schedule that is applied when the runtime schedule is used. 10 Format 11 C/C++ void omp_get_schedule(omp_sched_t * kind, int * modifier ); C/C++ 12 Fortran subroutine omp_get_schedule(kind, modifier) integer (kind=omp_sched_kind) kind integer modifier Fortran 13 14 Binding 15 The binding task set for an omp_get_schedule region is the generating task. Chapter 3 Runtime Library Routines 205 1 Effect 2 3 4 5 6 This routine returns the run-sched-var ICV in the task to which the routine binds. The first argument kind returns the schedule to be used. It can be any of the standard schedule types as defined in Section 3.2.12 on page 203, or any implementation specific schedule type. The second argument is interpreted as in the omp_set_schedule call, defined in Section 3.2.12 on page 203. 7 Cross References 8 • run-sched-var ICV, see Section 2.3 on page 34. 9 • omp_set_schedule routine, see Section 3.2.12 on page 203. 10 • OMP_SCHEDULE environment variable, see Section 4.1 on page 238. 11 • Determining the schedule of a worksharing loop, see Section 2.7.1.1 on page 59. 12 3.2.14 omp_get_thread_limit 13 Summary 14 15 The omp_get_thread_limit routine returns the maximum number of OpenMP threads available on the device. 16 Format 17 C/C++ int omp_get_thread_limit(void); C/C++ 18 Fortran integer function omp_get_thread_limit() Fortran 19 206 OpenMP API • Version 4.0 - July 2013 1 Binding 2 3 4 The binding thread set for an omp_get_thread_limit region is all threads on the device. The effect of executing this routine is not related to any specific region corresponding to any construct or API routine. 5 Effect 6 7 The omp_get_thread_limit routine returns the maximum number of OpenMP threads available on the device as stored in the ICV thread-limit-var. 8 Cross References 9 • thread-limit-var ICV, see Section 2.3 on page 34. • OMP_THREAD_LIMIT environment variable, see Section 4.10 on page 246. 10 11 3.2.15 omp_set_max_active_levels 12 Summary 13 14 The omp_set_max_active_levels routine limits the number of nested active parallel regions on the device, by setting the max-active-levels-var ICV. 15 Format 16 C/C++ void omp_set_max_active_levels (int max_levels); C/C++ 17 Chapter 3 Runtime Library Routines 207 1 Fortran subroutine omp_set_max_active_levels (max_levels) integer max_levels Fortran 2 3 Constraints on Arguments 4 5 The value of the argument passed to this routine must evaluate to a non-negative integer, otherwise the behavior of this routine is implementation defined. 6 Binding 7 8 9 10 When called from a sequential part of the program, the binding thread set for an omp_set_max_active_levels region is the encountering thread. When called from within any explicit parallel region, the binding thread set (and binding region, if required) for the omp_set_max_active_levels region is implementation defined. 11 Effect 12 13 The effect of this routine is to set the value of the max-active-levels-var ICV to the value specified in the argument. 14 15 16 If the number of parallel levels requested exceeds the number of levels of parallelism supported by the implementation, the value of the max-active-levels-var ICV will be set to the number of parallel levels supported by the implementation. 17 18 19 This routine has the described effect only when called from a sequential part of the program. When called from within an explicit parallel region, the effect of this routine is implementation defined. 20 Cross References 21 • max-active-levels-var ICV, see Section 2.3 on page 34. 22 • omp_get_max_active_levels routine, see Section 3.2.16 on page 209. 23 • OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.9 on page 245. 208 OpenMP API • Version 4.0 - July 2013 1 3.2.16 omp_get_max_active_levels 2 Summary 3 4 5 The omp_get_max_active_levels routine returns the value of the max-activelevels-var ICV, which determines the maximum number of nested active parallel regions on the device. 6 Format 7 C/C++ int omp_get_max_active_levels(void); C/C++ 8 Fortran integer function omp_get_max_active_levels() Fortran 9 10 Binding 11 12 13 14 When called from a sequential part of the program, the binding thread set for an omp_get_max_active_levels region is the encountering thread. When called from within any explicit parallel region, the binding thread set (and binding region, if required) for the omp_get_max_active_levels region is implementation defined. 15 Effect 16 17 18 The omp_get_max_active_levels routine returns the value of the max-activelevels-var ICV, which determines the maximum number of nested active parallel regions on the device. Chapter 3 Runtime Library Routines 209 1 Cross References 2 • max-active-levels-var ICV, see Section 2.3 on page 34. 3 • omp_set_max_active_levels routine, see Section 3.2.15 on page 207. 4 • OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.9 on page 245. 5 3.2.17 omp_get_level 6 Summary 7 The omp_get_level routine returns the value of the levels-var ICV. 8 Format 9 C/C++ int omp_get_level(void); C/C++ 10 Fortran integer function omp_get_level() Fortran 11 12 Binding 13 The binding task set for an omp_get_level region is the generating task. 210 OpenMP API • Version 4.0 - July 2013 1 Effect 2 3 4 5 The effect of the omp_get_level routine is to return the number of nested parallel regions (whether active or inactive) enclosing the current task such that all of the parallel regions are enclosed by the outermost initial task region on the current device. 6 Cross References 7 • levels-var ICV, see Section 2.3 on page 34. 8 • omp_get_active_level routine, see Section 3.2.20 on page 214. 9 • OMP_MAX_ACTIVE_LEVELS environment variable, see Section 4.9 on page 245. 10 3.2.18 omp_get_ancestor_thread_num 11 Summary 12 13 The omp_get_ancestor_thread_num routine returns, for a given nested level of the current thread, the thread number of the ancestor of the current thread. 14 Format 15 C/C++ int omp_get_ancestor_thread_num(int level); C/C++ 16 Fortran integer function omp_get_ancestor_thread_num(level) integer level Fortran 17 Chapter 3 Runtime Library Routines 211 1 Binding 2 3 4 The binding thread set for an omp_get_ancestor_thread_num region is the encountering thread. The binding region for an omp_get_ancestor_thread_num region is the innermost enclosing parallel region. 5 Effect 6 7 8 9 The omp_get_ancestor_thread_num routine returns the thread number of the ancestor at a given nest level of the current thread or the thread number of the current thread. If the requested nest level is outside the range of 0 and the nest level of the current thread, as returned by the omp_get_level routine, the routine returns -1. 10 11 12 Note – When the omp_get_ancestor_thread_num routine is called with a value of level=0, the routine always returns 0. If level=omp_get_level(), the routine has the same effect as the omp_get_thread_num routine. 13 Cross References 14 • omp_get_level routine, see Section 3.2.17 on page 210. 15 • omp_get_thread_num routine, see Section 3.2.4 on page 193. 16 • omp_get_team_size routine, see Section 3.2.19 on page 212. 17 3.2.19 omp_get_team_size 18 Summary 19 20 The omp_get_team_size routine returns, for a given nested level of the current thread, the size of the thread team to which the ancestor or the current thread belongs. 212 OpenMP API • Version 4.0 - July 2013 1 Format 2 C/C++ int omp_get_team_size(int level); C/C++ 3 Fortran integer function omp_get_team_size(level) integer level Fortran 4 5 Binding 6 7 8 The binding thread set for an omp_get_team_size region is the encountering thread. The binding region for an omp_get_team_size region is the innermost enclosing parallel region. 9 Effect 10 11 12 13 14 The omp_get_team_size routine returns the size of the thread team to which the ancestor or the current thread belongs. If the requested nested level is outside the range of 0 and the nested level of the current thread, as returned by the omp_get_level routine, the routine returns -1. Inactive parallel regions are regarded like active parallel regions executed with one thread. 15 16 17 Note – When the omp_get_team_size routine is called with a value of level=0, the routine always returns 1. If level=omp_get_level(), the routine has the same effect as the omp_get_num_threads routine. Chapter 3 Runtime Library Routines 213 1 Cross References 2 • omp_get_num_threads routine, see Section 3.2.2 on page 191. 3 • omp_get_level routine, see Section 3.2.17 on page 210. 4 • omp_get_ancestor_thread_num routine, see Section 3.2.18 on page 211. 5 3.2.20 omp_get_active_level 6 Summary 7 The omp_get_active_level routine returns the value of the active-level-vars ICV.. 8 Format 9 C/C++ int omp_get_active_level(void); C/C++ 10 Fortran integer function omp_get_active_level() Fortran 11 12 Binding 13 14 The binding task set for the an omp_get_active_level region is the generating task. 214 OpenMP API • Version 4.0 - July 2013 1 Effect 2 3 4 The effect of the omp_get_active_level routine is to return the number of nested, active parallel regions enclosing the current task such that all of the parallel regions are enclosed by the outermost initial task region on the current device. 5 Cross References 6 • active-levels-var ICV, see Section 2.3 on page 34. 7 • omp_get_level routine, see Section 3.2.17 on page 210. 8 9 3.2.21 omp_in_final Summary 10 11 The omp_in_final routine returns true if the routine is executed in a final task region; otherwise, it returns false. 12 Format 13 C/C++ int omp_in_final(void); C/C++ 14 Fortran logical function omp_in_final() Fortran 15 16 Binding 17 The binding task set for an omp_in_final region is the generating task. Chapter 3 Runtime Library Routines 215 1 Effect 2 3 omp_in_final returns true if the enclosing task region is final. Otherwise, it returns false. 4 3.2.22 omp_get_proc_bind 5 Summary 6 7 The omp_get_proc_bind routine returns the thread affinity policy to be used for the subsequent nested parallel regions that do not specify a proc_bind clause. 8 Format 9 C/C++ omp_proc_bind_t omp_get_proc_bind(void); C/C++ 10 Fortran 11 integer (kind=omp_proc_bind_kind) function omp_get_proc_bind() Fortran 12 13 Constraints on Arguments 14 15 16 17 The value returned by this routine must be one of the valid affinity policy kinds. 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: 216 OpenMP API • Version 4.0 - July 2013 1 C/C++ 2 3 4 5 6 7 8 typedef enum omp_proc_bind_t { omp_proc_bind_false = 0, omp_proc_bind_true = 1, omp_proc_bind_master = 2, omp_proc_bind_close = 3, omp_proc_bind_spread = 4 } omp_proc_bind_t; C/C++ 9 Fortran 10 11 integer (kind=omp_proc_bind_kind), & parameter :: omp_proc_bind_false = 0 12 13 integer (kind=omp_proc_bind_kind), & parameter :: omp_proc_bind_true = 1 14 15 integer (kind=omp_proc_bind_kind), & parameter :: omp_proc_bind_master = 2 16 17 integer (kind=omp_proc_bind_kind), & parameter :: omp_proc_bind_close = 3 18 19 integer (kind=omp_proc_bind_kind), & parameter :: omp_proc_bind_spread = 4 Fortran 20 Binding 21 The binding task set for an omp_get_proc_bind region is the generating task. 22 Effect 23 24 25 The effect of this routine is to return the value of the first element of the bind-var ICV of the current task. See Section 2.5.2 on page 49 for the rules governing the thread affinity policy. 26 Cross References 27 • bind-var ICV, see Section 2.3 on page 34. 28 • OMP_PROC_BIND environment variable, see Section 4.4 on page 241. 29 • Controlling OpenMP thread affinity, see Section 2.5.2 on page 49. Chapter 3 Runtime Library Routines 217 1 3.2.23 omp_set_default_device 2 Summary 3 4 The omp_set_default_device routine controls the default target device by assigning the value of the default-device-var ICV. 5 Format 6 C/C++ void omp_set_default_device(int device_num ); C/C++ 7 Fortran subroutine omp_set_default_device( device_num ) integer device_num Fortran Binding 8 9 10 The binding task set for an omp_set_default_device region is the generating task. 11 Effect 12 13 14 The effect of this routine is to set the value of the default-device-var ICV of the current task to the value specified in the argument. When called from within a target region the effect of this routine is unspecified. 15 Cross References: 16 • default-device-var, see Section 2.3 on page 34. 17 • omp_get_default_device, see Section 3.2.24 on page 219. 18 • OMP_DEFAULT_DEVICE environment variable, see Section 4.13 on page 248 218 OpenMP API • Version 4.0 - July 2013 1 3.2.24 omp_get_default_device 2 Summary 3 The omp_get_default_device routine returns the default target device. 4 Format C/C++ int omp_get_default_device(void); C/C++ 5 Fortran integer function omp_get_default_device() Fortran 6 7 Binding 8 9 The binding task set for an omp_get_default_device region is the generating task. 10 Effect 11 12 13 The omp_get_default_device routine returns the value of the default-device-var ICV of the current task. When called from within a target region the effect of this routine is unspecified. 14 Cross References 15 • default-device-var, see Section 2.3 on page 34. 16 • omp_set_default_device, see Section 3.2.23 on page 218. 17 • OMP_DEFAULT_DEVICE environment variable, see Section 4.13 on page 248. Chapter 3 Runtime Library Routines 219 1 3.2.25 omp_get_num_devices 2 Summary 3 The omp_get_num_devices routine returns the number of target devices. 4 Format C/C++ 5 int omp_get_num_devices(void); C/C++ 6 Fortran integer function omp_get_num_devices() Fortran 7 8 Binding 9 The binding task set for an omp_get_num_devices region is the generating task. 10 Effect 11 12 The omp_get_num_devices routine returns the number of available target devices. When called from within a target region the effect of this routine is unspecified. 13 Cross References: 14 None. 220 OpenMP API • Version 4.0 - July 2013 1 3.2.26 omp_get_num_teams 2 Summary 3 4 The omp_get_num_teams routine returns the number of teams in the current teams region. 5 Format C/C++ 6 int omp_get_num_teams(void); C/C++ 7 Fortran integer function omp_get_num_teams() Fortran 8 9 Binding 10 The binding task set for an omp_get_num_teams region is the generating task. 11 Effect 12 13 The effect of this routine is to return the number of teams in the current teams region. The routine returns 1 if it is called from outside of a teams region. 14 Cross References: 15 • teams construct, see Section 2.9.5 on page 86. Chapter 3 Runtime Library Routines 221 1 3.2.27 omp_get_team_num 2 Summary 3 The omp_get_team_num routine returns the team number of the calling thread. 4 Format C/C++ 5 int omp_get_team_num(void); C/C++ 6 Fortran integer function omp_get_team_num() Fortran 7 8 Binding 9 The binding task set for an omp_get_team_num region is the generating task. 10 Effect 11 12 13 14 The omp_get_team_num routine returns the team number of the calling thread. The team number is an integer between 0 and one less than the value returned by omp_get_num_teams, inclusive. The routine returns 0 if it is called outside of a teams region. 15 Cross References: 16 • teams construct, see Section 2.9.5 on page 86. 17 • omp_get_num_teams routine, see Section 3.2.26 on page 221. 222 OpenMP API • Version 4.0 - July 2013 1 3.2.28 omp_is_initial_device 2 Summary 3 4 The omp_is_initial_device routine returns true if the current task is executing on the host device; otherwise, it returns false. 5 Format 6 C/C++ int omp_is_initial_device(void); C/C++ 7 8 Fortran logical function omp_is_initial_device() Fortran 9 10 Binding 11 The binding task set for an omp_is_initial_device region is the generating task. 12 Effect 13 14 The effect of this routine is to return true if the current task is executing on the host device; otherwise, it returns false. 15 Cross References: 16 • target construct, see Section 2.9.2 on page 79. Chapter 3 Runtime Library Routines 223 1 3.3 Lock Routines 2 3 4 5 6 The OpenMP runtime library includes a set of general-purpose lock routines that can be used for synchronization. These general-purpose lock routines operate on OpenMP locks that are represented by OpenMP lock variables. OpenMP lock variables must be accessed only through the routines described in this section; programs that otherwise access OpenMP lock variables are non-conforming. 7 8 9 10 11 An OpenMP lock can be in one of the following states: uninitialized, unlocked, or locked. If a lock is in the unlocked state, a task can set the lock, which changes its state to locked. The task that sets the lock is then said to own the lock. A task that owns a lock can unset that lock, returning it to the unlocked state. A program in which a task unsets a lock that is owned by another task is non-conforming. 12 13 14 15 16 Two types of locks are supported: simple locks and nestable locks. A nestable lock can be set multiple times by the same task before being unset; a simple lock cannot be set if it is already owned by the task trying to set it. Simple lock variables are associated with simple locks and can only be passed to simple lock routines. Nestable lock variables are associated with nestable locks and can only be passed to nestable lock routines. 17 18 19 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. 20 21 22 23 The OpenMP lock routines access a lock variable in such a way that they always read and update the most current value of the lock variable. It is not necessary for an OpenMP program to include explicit flush directives to ensure that the lock variable’s value is consistent among different tasks. 24 Binding 25 26 27 28 The binding thread set for all lock routine regions is all threads in the contention group. As a consequence, for each OpenMP lock, the lock routine effects relate to all tasks that call the routines, without regard to which teams the threads in the contention group executing the tasks belong. 224 OpenMP API • Version 4.0 - July 2013 1 Simple Lock Routines 2 3 4 The type omp_lock_t is a data type capable of representing a simple lock. For the following routines, a simple lock variable must be of omp_lock_t type. All simple lock routines require an argument that is a pointer to a variable of type omp_lock_t. C/C++ C/C++ Fortran 5 6 For the following routines, a simple lock variable must be an integer variable of kind=omp_lock_kind. Fortran 7 The simple lock routines are as follows: 8 • The omp_init_lock routine initializes a simple lock. 9 • The omp_destroy_lock routine uninitializes a simple lock. 10 • The omp_set_lock routine waits until a simple lock is available, and then sets it. 11 • The omp_unset_lock routine unsets a simple lock. 12 • The omp_test_lock routine tests a simple lock, and sets it if it is available. 13 Nestable Lock Routines: 14 15 16 17 The type omp_nest_lock_t is a data type capable of representing a nestable lock. For the following routines, a nested lock variable must be of omp_nest_lock_t type. All nestable lock routines require an argument that is a pointer to a variable of type omp_nest_lock_t. C/C++ C/C++ Fortran 18 19 For the following routines, a nested lock variable must be an integer variable of kind=omp_nest_lock_kind. Fortran 20 The nestable lock routines are as follows: 21 • The omp_init_nest_lock routine initializes a nestable lock. 22 • The omp_destroy_nest_lock routine uninitializes a nestable lock. 23 24 • The omp_set_nest_lock routine waits until a nestable lock is available, and then sets it. Chapter 3 Runtime Library Routines 225 1 • The omp_unset_nest_lock routine unsets a nestable lock. 2 3 • The omp_test_nest_lock routine tests a nestable lock, and sets it if it is 4 Restrictions 5 OpenMP lock routines have the following restrictions: 6 7 • The use of the same OpenMP lock in different contention groups results in 8 available. unspecified behavior. 3.3.1 omp_init_lock and omp_init_nest_lock Summary 9 10 These routines provide the only means of initializing an OpenMP lock. 11 Format C/C++ void omp_init_lock(omp_lock_t *lock); void omp_init_nest_lock(omp_nest_lock_t *lock); C/C++ 12 Fortran subroutine omp_init_lock(svar) integer (kind=omp_lock_kind) svar subroutine omp_init_nest_lock(nvar) integer (kind=omp_nest_lock_kind) nvar Fortran 13 226 OpenMP API • Version 4.0 - July 2013 1 Constraints on Arguments 2 3 A program that accesses a lock that is not in the uninitialized state through either routine is non-conforming. 4 Effect 5 6 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. 8 omp_destroy_lock and omp_destroy_nest_lock 9 Summary 7 3.3.2 10 These routines ensure that the OpenMP lock is uninitialized. 11 Format C/C++ void omp_destroy_lock(omp_lock_t *lock); void omp_destroy_nest_lock(omp_nest_lock_t *lock); 12 C/C++ Fortran subroutine omp_destroy_lock(svar) integer (kind=omp_lock_kind) svar subroutine omp_destroy_nest_lock(nvar) integer (kind=omp_nest_lock_kind) nvar 13 Fortran Chapter 3 Runtime Library Routines 227 1 Constraints on Arguments 2 3 A program that accesses a lock that is not in the unlocked state through either routine is non-conforming. 4 Effect 5 The effect of these routines is to change the state of the lock to uninitialized. 6 3.3.3 omp_set_lock and omp_set_nest_lock 7 Summary 8 9 These routines provide a means of setting an OpenMP lock. The calling task region is suspended until the lock is set. Format 10 C/C++ void omp_set_lock(omp_lock_t *lock); void omp_set_nest_lock(omp_nest_lock_t *lock); C/C++ 11 Fortran subroutine omp_set_lock(svar) integer (kind=omp_lock_kind) svar subroutine omp_set_nest_lock(nvar) integer (kind=omp_nest_lock_kind) nvar Fortran 12 228 OpenMP API • Version 4.0 - July 2013 1 Constraints on Arguments 2 3 4 A program that accesses a lock that is in the uninitialized state through either routine is non-conforming. 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. 5 Effect 6 7 Each of these routines causes suspension of the task executing the routine until the specified lock is available and then sets the lock. 8 9 A simple lock is available if it is unlocked. Ownership of the lock is granted to the task executing the routine. 10 11 12 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. 13 14 3.3.4 omp_unset_lock and omp_unset_nest_lock 15 Summary 16 These routines provide the means of unsetting an OpenMP lock. Chapter 3 Runtime Library Routines 229 Format 1 C/C++ void omp_unset_lock(omp_lock_t *lock); void omp_unset_nest_lock(omp_nest_lock_t *lock); C/C++ 2 Fortran subroutine omp_unset_lock(svar) integer (kind=omp_lock_kind) svar subroutine omp_unset_nest_lock(nvar) integer (kind=omp_nest_lock_kind) nvar Fortran 3 4 Constraints on Arguments 5 6 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. 7 Effect 8 For a simple lock, the omp_unset_lock routine causes the lock to become unlocked. 9 10 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. 11 12 13 For either routine, if the lock becomes unlocked, and if one or more task regions were suspended because the lock was unavailable, the effect is that one task is chosen and given ownership of the lock. 230 OpenMP API • Version 4.0 - July 2013 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++ int omp_test_lock(omp_lock_t *lock); int omp_test_nest_lock(omp_nest_lock_t *lock); C/C++ 6 Fortran logical integer integer integer function omp_test_lock(svar) (kind=omp_lock_kind) svar function omp_test_nest_lock(nvar) (kind=omp_nest_lock_kind) nvar Fortran 7 8 Constraints on Arguments 9 10 11 A program that accesses a lock that is in the uninitialized state through either routine is non-conforming. The behavior is unspecified if a simple lock accessed by omp_test_lock is in the locked state and is owned by the task that contains the call. 12 Effect 13 14 15 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. 16 17 For a simple lock, the omp_test_lock routine returns true if the lock is successfully set; otherwise, it returns false. Chapter 3 Runtime Library Routines 231 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 2 232 OpenMP API • Version 4.0 - July 2013 1 3.4 This section describes routines that support a portable wall clock timer. 2 3 Timing Routines 3.4.1 omp_get_wtime 4 Summary 5 The omp_get_wtime routine returns elapsed wall clock time in seconds. 6 Format C/C++ double omp_get_wtime(void); C/C++ 7 Fortran double precision function omp_get_wtime() Fortran 8 9 Binding 10 11 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. 12 Effect 13 14 15 16 17 The omp_get_wtime routine returns a value equal to the elapsed wall clock time in seconds since some “time in the past”. The actual “time in the past” is arbitrary, but it is guaranteed not to change during the execution of the application program. The time returned is a “per-thread time”, so it is not required to be globally consistent across all the threads participating in an application. Chapter 3 Runtime Library Routines 233 Note – It is anticipated that the routine will be used to measure elapsed times as shown in the following example: 1 2 C/C++ double start; double end; start = omp_get_wtime(); ... work to be timed ... end = omp_get_wtime(); printf("Work took %f seconds\n", end - start); C/C++ 3 Fortran 4 DOUBLE PRECISION START, END START = omp_get_wtime() ... work to be timed ... END = omp_get_wtime() PRINT *, "Work took", END - START, "seconds" Fortran 5 6 3.4.2 omp_get_wtick 7 Summary 8 9 The omp_get_wtick routine returns the precision of the timer used by omp_get_wtime. 234 OpenMP API • Version 4.0 - July 2013 1 Format C/C++ double omp_get_wtick(void); C/C++ 2 Fortran double precision function omp_get_wtick() Fortran 3 4 Binding 5 6 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. 7 Effect 8 9 The omp_get_wtick routine returns a value equal to the number of seconds between successive clock ticks of the timer used by omp_get_wtime. Chapter 3 Runtime Library Routines 235 236 OpenMP API • Version 4.0 - July 2013 1 2 CHAPTER 4 Environment Variables 3 4 5 6 7 8 9 10 This chapter describes the OpenMP environment variables that specify the settings of the ICVs that affect the execution of OpenMP programs (see Section 2.3 on page 34). The names of the environment variables must be upper case. The values assigned to the environment variables are case insensitive and may have leading and trailing white space. Modifications to the environment variables after the program has started, even if modified by the program itself, are ignored by the OpenMP implementation. However, the settings of some of the ICVs can be modified during the execution of the OpenMP program by the use of the appropriate directive clauses or OpenMP API routines. 11 The environment variables are as follows: 12 13 • OMP_SCHEDULE sets the run-sched-var ICV that specifies the runtime schedule type 14 15 • OMP_NUM_THREADS sets the nthreads-var ICV that specifies the number of threads 16 17 • OMP_DYNAMIC sets the dyn-var ICV that specifies the dynamic adjustment of 18 19 • OMP_PROC_BIND sets the bind-var ICV that controls the OpenMP thread affinity 20 21 • OMP_PLACES sets the place-partition-var ICV that defines the OpenMP places that 22 • OMP_NESTED sets the nest-var ICV that enables or disables nested parallelism. 23 24 • OMP_STACKSIZE sets the stacksize-var ICV that specifies the size of the stack for 25 26 • OMP_WAIT_POLICY sets the wait-policy-var ICV that controls the desired behavior 27 28 • OMP_MAX_ACTIVE_LEVELS sets the max-active-levels-var ICV that controls the 29 30 • OMP_THREAD_LIMIT sets the thread-limit-var ICV that controls the maximum and chunk size. It can be set to any of the valid OpenMP schedule types. to use for parallel regions. threads to use for parallel regions. policy. are available to the execution environment. threads created by the OpenMP implementation. of waiting threads. maximum number of nested active parallel regions. number of threads participating in the OpenMP program. 237 1 • OMP_CANCELLATION sets the cancel-var ICV that enables or disables cancellation. 2 3 • OMP_DISPLAY_ENV instructs the runtime to display the OpenMP version number 4 5 • OMP_DEFAULT_DEVICE sets the default-device-var ICV that controls the default 6 7 8 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: 9 • csh: and the initial values of the ICVs, once, during initialization of the runtime. device number. setenv OMP_SCHEDULE "dynamic" • ksh: 10 export OMP_SCHEDULE="dynamic" • DOS: 11 set OMP_SCHEDULE=dynamic 12 4.1 OMP_SCHEDULE 13 14 15 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. 16 The value of this environment variable takes the form: 17 type[,chunk] 18 where 19 • type is one of static, dynamic, guided, or auto 20 • chunk is an optional positive integer that specifies the chunk size 21 22 If chunk is present, there may be white space on either side of the “,”. See Section 2.7.1 on page 53 for a detailed description of the schedule types. 23 24 The behavior of the program is implementation defined if the value of OMP_SCHEDULE does not conform to the above format. 238 OpenMP API • Version 4.0 - July 2013 1 2 3 Implementation specific schedules cannot be specified in OMP_SCHEDULE. They can only be specified by calling omp_set_schedule, described in Section 3.2.12 on page 203. 4 Example: setenv OMP_SCHEDULE "guided,4" setenv OMP_SCHEDULE "dynamic" 5 Cross References 6 • run-sched-var ICV, see Section 2.3 on page 34. 7 • Loop construct, see Section 2.7.1 on page 53. 8 • Parallel loop construct, see Section 2.10.1 on page 95. 9 • omp_set_schedule routine, see Section 3.2.12 on page 203. 10 • omp_get_schedule routine, see Section 3.2.13 on page 205. 11 4.2 OMP_NUM_THREADS 12 13 14 15 16 17 18 The OMP_NUM_THREADS environment variable sets the number of threads to use for parallel regions by setting the initial value of the nthreads-var ICV. See Section 2.3 on page 34 for a comprehensive set of rules about the interaction between the OMP_NUM_THREADS environment variable, the num_threads clause, the omp_set_num_threads library routine and dynamic adjustment of threads, and Section 2.5.1 on page 47 for a complete algorithm that describes how the number of threads for a parallel region is determined. 19 20 21 The value of this environment variable must be a list of positive integer values. The values of the list set the number of threads to use for parallel regions at the corresponding nested levels. 22 23 24 The behavior of the program is implementation defined if any value of the list specified in the OMP_NUM_THREADS environment variable leads to a number of threads which is greater than an implementation can support, or if any value is not a positive integer. 25 Example: setenv OMP_NUM_THREADS 4,3,2 Chapter 4 Environment Variables 239 1 Cross References: 2 • nthreads-var ICV, see Section 2.3 on page 34. 3 • num_threads clause, Section 2.5 on page 44. 4 • omp_set_num_threads routine, see Section 3.2.1 on page 189. 5 • omp_get_num_threads routine, see Section 3.2.2 on page 191. 6 • omp_get_max_threads routine, see Section 3.2.3 on page 192. 7 • omp_get_team_size routine, see Section 3.2.19 on page 212. 8 9 4.3 OMP_DYNAMIC 10 11 12 13 14 15 16 17 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. 18 Example: setenv OMP_DYNAMIC true 19 Cross References: 20 • dyn-var ICV, see Section 2.3 on page 34. 21 • omp_set_dynamic routine, see Section 3.2.7 on page 197. 22 • omp_get_dynamic routine, see Section 3.2.8 on page 198. 240 OpenMP API • Version 4.0 - July 2013 1 4.4 OMP_PROC_BIND 2 3 4 5 The OMP_PROC_BIND environment variable sets the initial value of the bind-var ICV. The value of this environment variable is either true, false, or a comma separated list of master, close, or spread. The values of the list set the thread affinity policy to be used for parallel regions at the corresponding nested level. 6 7 8 If the environment variable is set to false, the execution environment may move OpenMP threads between OpenMP places, thread affinity is disabled, and proc_bind clauses on parallel constructs are ignored. 9 10 11 Otherwise, the execution environment should not move OpenMP threads between OpenMP places, thread affinity is enabled, and the initial thread is bound to the first place in the OpenMP place list. 12 13 14 15 The behavior of the program is implementation defined if any of the values in the OMP_PROC_BIND environment variable is not true, false, or a comma separated list of master, close, or spread. The behavior is also implementation defined if an initial thread cannot be bound to the first place in the OpenMP place list. 16 Example: setenv OMP_PROC_BIND false setenv OMP_PROC_BIND "spread, spread, close" 17 Cross References: 18 • bind-var ICV, see Section 2.3 on page 34. 19 • proc_bind clause, see Section 2.5.2 on page 49. 20 • omp_get_proc_bind routine, see Section 3.2.22 on page 216. 21 22 23 24 25 26 4.5 OMP_PLACES A list of places can be specified in the OMP_PLACES environment variable. The placepartition-var ICV obtains its initial value from the OMP_PLACES value, and makes the list available to the execution environment. The value of OMP_PLACES can be one of two types of values: either an abstract name describing a set of places or an explicit list of places described by non-negative numbers. Chapter 4 Environment Variables 241 1 2 3 4 5 6 The OMP_PLACES environment variable can be defined using an explicit ordered list of comma-separated places. A place is defined by an unordered set of comma-separated non-negative numbers enclosed by braces. The meaning of the numbers and how the numbering is done are implementation defined. Generally, the numbers represent the smallest unit of execution exposed by the execution environment, typically a hardware thread. 7 8 9 10 11 Intervals may also be used to define places. Intervals can be specified using the <lowerbound> : <length> : <stride> notation to represent the following list of numbers: “<lower-bound>, <lower-bound> + <stride>, …, <lower-bound> + (<length>1)*<stride>.” When <stride> is omitted, a unit stride is assumed. Intervals can specify numbers within a place as well as sequences of places. 12 13 An exclusion operator “!” can also be used to exclude the number or place immediately following the operator. 14 15 16 17 Alternatively, the abstract names listed in TABLE 4-1 should be understood by the execution and runtime environment. The precise definitions of the abstract names are implementation defined. An implementation may also add abstract names as appropriate for the target platform. 18 19 20 21 22 23 The abstract name may be appended by a positive number in parentheses to denote the length of the place list to be created, that is abstract_name(num-places). When requesting fewer places than available on the system, the determination of which resources of type abstract_name are to be included in the place list is implementation defined. When requesting more resources than available, the length of the place list is implementation defined. TABLE 4-1 List of defined abstract names for OMP_PLACES Abstract Name Meaning threads Each place corresponds to a single hardware thread on the target machine. cores Each place corresponds to a single core (having one or more hardware threads) on the target machine. sockets Each place corresponds to a single socket (consisting of one or more cores) on the target machine. The behavior of the program is implementation defined when the execution environment cannot map a numerical value (either explicitly defined or implicitly derived from an interval) within the OMP_PLACES list to a processor on the target platform, or if it maps to an unavailable processor. The behavior is also implementation defined when the OMP_PLACES environment variable is defined using an abstract name. 24 25 26 27 28 242 OpenMP API • Version 4.0 - July 2013 Example: 1 setenv setenv setenv setenv setenv OMP_PLACES OMP_PLACES OMP_PLACES OMP_PLACES OMP_PLACES threads "threads(4)" "{0,1,2,3},{4,5,6,7},{8,9,10,11},{12,13,14,15}" "{0:4},{4:4},{8:4},{12:4}" "{0:4}:4:4" 2 3 4 where each of the last three definitions corresponds to the same 4 places including the smallest units of execution exposed by the execution environment numbered, in turn, 0 to 3, 4 to 7, 8 to 11, and 12 to 15. 5 Cross References 6 • place-partition-var, Section 2.3 on page 34. 7 • Controlling OpenMP thread affinity, Section 2.5.2 on page 49. 8 4.6 OMP_NESTED 9 10 11 12 13 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. 14 Example: setenv OMP_NESTED false 15 Cross References 16 • nest-var ICV, see Section 2.3 on page 34. 17 • omp_set_nested routine, see Section 3.2.10 on page 200. 18 • omp_get_nested routine, see Section 3.2.19 on page 212. 19 Chapter 4 Environment Variables 243 1 4.7 OMP_STACKSIZE 2 3 4 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 an initial thread. 5 The value of this environment variable takes the form: 6 size | sizeB | sizeK | sizeM | sizeG 7 where: 8 9 • size is a positive integer that specifies the size of the stack for threads that are created by the OpenMP implementation. 10 11 12 13 • B, K, M, and G are letters that specify whether the given size is in Bytes, Kilobytes 14 15 If only size is specified and none of B, K, M, or G is specified, then size is assumed to be in Kilobytes. 16 17 18 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. 19 Examples: (1024 Bytes), Megabytes (1024 Kilobytes), or Gigabytes (1024 Megabytes), respectively. If one of these letters is present, there may be white space between size and the letter. setenv setenv setenv setenv setenv setenv setenv OMP_STACKSIZE OMP_STACKSIZE OMP_STACKSIZE OMP_STACKSIZE OMP_STACKSIZE OMP_STACKSIZE OMP_STACKSIZE 2000500B "3000 k " 10M " 10 M " "20 m " " 1G" 20000 20 Cross References 21 • stacksize-var ICV, see Section 2.3 on page 34. 244 OpenMP API • Version 4.0 - July 2013 1 4.8 OMP_WAIT_POLICY 2 3 4 5 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. 6 The value of this environment variable takes the form: 7 ACTIVE | PASSIVE 8 9 10 The ACTIVE value specifies that waiting threads should mostly be active, consuming processor cycles, while waiting. An OpenMP implementation may, for example, make waiting threads spin. 11 12 13 The PASSIVE value specifies that waiting threads should mostly be passive, not consuming processor cycles, while waiting. For example, an OpenMP implementation may make waiting threads yield the processor to other threads or go to sleep. 14 The details of the ACTIVE and PASSIVE behaviors are implementation defined. 15 Examples: setenv setenv setenv setenv OMP_WAIT_POLICY OMP_WAIT_POLICY OMP_WAIT_POLICY OMP_WAIT_POLICY ACTIVE active PASSIVE passive 16 Cross References 17 • wait-policy-var ICV, see Section 2.3 on page 24. 18 19 20 21 4.9 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-levelsvar ICV. Chapter 4 Environment Variables 245 1 2 3 4 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 34. 8 • omp_set_max_active_levels routine, see Section 3.2.15 on page 207. 9 • omp_get_max_active_levels routine, see Section 3.2.16 on page 209. 10 4.10 OMP_THREAD_LIMIT 11 12 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. 13 14 15 16 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. 17 Cross References 18 • thread-limit-var ICV, see Section 2.3 on page 34. 19 • omp_get_thread_limit routine, see Section 3.2.14 on page 206. 20 4.11 OMP_CANCELLATION 21 22 The OMP_CANCELLATION environment variable sets the initial value of the cancel-var ICV. 23 24 25 26 The value of this environment variable must be true or false. If set to true, the effects of the cancel construct and of cancellation points are enabled and cancellation is activated. If set to false, cancellation is disabled and the cancel construct and cancellation points are effectively ignored. 246 OpenMP API • Version 4.0 - July 2013 1 Cross References: 2 • cancel-var, see Section 2.3.1 on page 35. 3 • cancel construct, see Section 2.13.1 on page 140. 4 • cancellation point construct, see Section 2.13.2 on page 143 5 • omp_get_cancellation routine, see Section 3.2.9 on page 199 6 4.12 OMP_DISPLAY_ENV 7 8 9 10 11 The OMP_DISPLAY_ENV environment variable instructs the runtime to display the OpenMP version number and the value of the ICVs associated with the environment variables described in Chapter 4, as name = value pairs. The runtime displays this information once, after processing the environment variables and before any user calls to change the ICV values by runtime routines defined in Chapter 3. 12 13 The value of the OMP_DISPLAY_ENV environment variable may be set to one of these values: 14 TRUE | FALSE | VERBOSE 15 16 17 18 19 20 21 The TRUE value instructs the runtime to display the OpenMP version number defined by the _OPENMP version macro (or the openmp_version Fortran parameter) value and the initial ICV values for the environment variables listed in Chapter 4. The VERBOSE value indicates that the runtime may also display the values of runtime variables that may be modified by vendor-specific environment variables. The runtime does not display any information when the OMP_DISPLAY_ENV environment variable is FALSE, undefined, or any other value than TRUE or VERBOSE. 22 23 24 25 26 27 28 The display begins with "OPENMP DISPLAY ENVIRONMENT BEGIN", followed by the _OPENMP version macro (or the openmp_version Fortran parameter) value and ICV values, in the format NAME '=' VALUE. NAME corresponds to the macro or environment variable name, optionally prepended by a bracketed device-type. VALUE corresponds to the value of the macro or ICV associated with this environment variable. Values should be enclosed in single quotes. The display is terminated with "OPENMP DISPLAY ENVIRONMENT END". 29 30 Example: % setenv OMP_DISPLAY_ENV TRUE Chapter 4 Environment Variables 247 The above example causes an OpenMP implementation to generate output of the following form: 1 2 OPENMP DISPLAY ENVIRONMENT BEGIN _OPENMP='201307' [host] OMP_SCHEDULE='GUIDED,4' [host] OMP_NUM_THREADS='4,3,2' [device] OMP_NUM_THREADS='2' [host,device] OMP_DYNAMIC='TRUE' [host] OMP_PLACES='{0:4},{4:4},{8:4},{12:4}' ... OPENMP DISPLAY ENVIRONMENT END 3 4.13 OMP_DEFAULT_DEVICE 4 5 The OMP_DEFAULT_DEVICE environment variable sets the device number to use in device constructs by setting the initial value of the default-device-var ICV. 6 The value of this environment variable must be a non-negative integer value. 7 Cross References: 8 • default-device-var ICV, see Section 2.3 on page 34. 9 • device constructs, Section 2.9 on page 77 248 OpenMP API • Version 4.0 - July 2013 1 APPENDIX A 3 Stubs for Runtime Library Routines 4 5 6 7 8 This section provides stubs for the runtime library routines defined in the OpenMP API. The stubs are provided to enable portability to platforms that do not support the OpenMP API. On these platforms, OpenMP programs must be linked with a library containing these stub routines. The stub routines assume that the directives in the OpenMP program are ignored. As such, they emulate serial semantics. 9 10 11 Note that the lock variable that appears in the lock routines must be accessed exclusively through these routines. It should not be initialized or otherwise modified in the user program. 12 13 14 15 In an actual implementation the lock variable might be used to hold the address of an allocated memory block, but here it is used to hold an integer value. Users should not make assumptions about mechanisms used by OpenMP implementations to implement locks based on the scheme used by the stub procedures. 2 Fortran 16 17 18 19 Note – In order to be able to compile the Fortran stubs file, the include file omp_lib.h was split into two files: omp_lib_kinds.h and omp_lib.h and the omp_lib_kinds.h file included where needed. There is no requirement for the implementation to provide separate files. Fortran 249 1 A.1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 C/C++ Stub Routines #include <stdio.h> #include <stdlib.h> #include "omp.h" void omp_set_num_threads(int num_threads) { } int omp_get_num_threads(void) { return 1; } int omp_get_max_threads(void) { return 1; } int omp_get_thread_num(void) { return 0; } int omp_get_num_procs(void) { return 1; } int omp_in_parallel(void) { return 0; } void omp_set_dynamic(int dynamic_threads) { } int omp_get_dynamic(void) { return 0; } int omp_get_cancellation(void) { return 0; } 250 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 void omp_set_nested(int nested) { } int omp_get_nested(void) { return 0; } void omp_set_schedule(omp_sched_t kind, int modifier) { } void omp_get_schedule(omp_sched_t *kind, int *modifier) { *kind = omp_sched_static; *modifier = 0; } int omp_get_thread_limit(void) { return 1; } void omp_set_max_active_levels(int max_active_levels) { } int omp_get_max_active_levels(void) { return 0; } int omp_get_level(void) { return 0; } int omp_get_ancestor_thread_num(int level) { if (level == 0) { return 0; } else { return -1; } } Appendix A Stubs for Runtime Library Routines 251 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 int omp_get_team_size(int level) { if (level == 0) { return 1; } else { return -1; } } int omp_get_active_level(void) { return 0; } int omp_in_final(void) { return 1; } omp_proc_bind_t omp_get_proc_bind(void) { return omp_proc_bind_false; } void omp_set_default_device(int device_num) { } int omp_get_default_device(void) { return 0; } int omp_get_num_devices(void) { return 0; } int omp_get_num_teams(void) { return 1; } int omp_get_team_num(void) { return 0; } 252 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 int omp_is_initial_device(void) { return 1; } struct __omp_lock { int lock; }; enum { UNLOCKED = -1, INIT, LOCKED }; void omp_init_lock(omp_lock_t *arg) { struct __omp_lock *lock = (struct __omp_lock *)arg; lock->lock = UNLOCKED; } void omp_destroy_lock(omp_lock_t *arg) { struct __omp_lock *lock = (struct __omp_lock *)arg; lock->lock = INIT; } void omp_set_lock(omp_lock_t *arg) { struct __omp_lock *lock = (struct __omp_lock *)arg; if (lock->lock == UNLOCKED) { lock->lock = LOCKED; } else if (lock->lock == LOCKED) { fprintf(stderr, "error: deadlock in using lock variable\n"); exit(1); } else { fprintf(stderr, "error: lock not initialized\n"); exit(1); } } void omp_unset_lock(omp_lock_t *arg) { struct __omp_lock *lock = (struct __omp_lock *)arg; if (lock->lock == LOCKED) { lock->lock = UNLOCKED; } else if (lock->lock == UNLOCKED) { Appendix A Stubs for Runtime Library Routines 253 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 fprintf(stderr, "error: lock not set\n"); exit(1); } else { fprintf(stderr, "error: lock not initialized\n"); exit(1); } } int omp_test_lock(omp_lock_t *arg) { struct __omp_lock *lock = (struct __omp_lock *)arg; if (lock->lock == UNLOCKED) { lock->lock = LOCKED; return 1; } else if (lock->lock == LOCKED) { return 0; } else { fprintf(stderr, "error: lock not initialized\n"); exit(1); } } struct __omp_nest_lock { short owner; short count; }; enum { NOOWNER = -1, MASTER = 0 }; void omp_init_nest_lock(omp_nest_lock_t *arg) { struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg; nlock->owner = NOOWNER; nlock->count = 0; } void omp_destroy_nest_lock(omp_nest_lock_t *arg) { struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg; nlock->owner = NOOWNER; nlock->count = UNLOCKED; } 254 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 void omp_set_nest_lock(omp_nest_lock_t *arg) { struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg; if (nlock->owner == MASTER && nlock->count >= 1) { nlock->count++; } else if (nlock->owner == NOOWNER && nlock->count == 0) { nlock->owner = MASTER; nlock->count = 1; } else { fprintf(stderr, "error: lock corrupted or not initialized\n"); exit(1); } } void omp_unset_nest_lock(omp_nest_lock_t *arg) { struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg; if (nlock->owner == MASTER && nlock->count >= 1) { nlock->count--; if (nlock->count == 0) { nlock->owner = NOOWNER; } } else if (nlock->owner == NOOWNER && nlock->count == 0) { fprintf(stderr, "error: lock not set\n"); exit(1); } else { fprintf(stderr, "error: lock corrupted or not initialized\n"); exit(1); } } int omp_test_nest_lock(omp_nest_lock_t *arg) { struct __omp_nest_lock *nlock=(struct __omp_nest_lock *)arg; omp_set_nest_lock(arg); return nlock->count; } Appendix A Stubs for Runtime Library Routines 255 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 double omp_get_wtime(void) { /* This function does not provide a working * wallclock timer. Replace it with a version * customized for the target machine. */ return 0.0; } double omp_get_wtick(void) { /* This function does not provide a working * clock tick function. Replace it with * a version customized for the target machine. */ return 365. * 86400.; } 18 256 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 A.2 Fortran Stub Routines subroutine omp_set_num_threads(num_threads) integer num_threads return end subroutine integer function omp_get_num_threads() omp_get_num_threads = 1 return end function integer function omp_get_max_threads() omp_get_max_threads = 1 return end function integer function omp_get_thread_num() omp_get_thread_num = 0 return end function integer function omp_get_num_procs() omp_get_num_procs = 1 return end function logical function omp_in_parallel() omp_in_parallel = .false. return end function subroutine omp_set_dynamic(dynamic_threads) logical dynamic_threads return end subroutine logical function omp_get_dynamic() omp_get_dynamic = .false. return end function logical function omp_get_cancellation() omp_get_cancellation = .false. return end function Appendix A Stubs for Runtime Library Routines 257 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 subroutine omp_set_nested(nested) logical nested return end subroutine logical function omp_get_nested() omp_get_nested = .false. return end function subroutine omp_set_schedule(kind, modifier) include 'omp_lib_kinds.h' integer (kind=omp_sched_kind) kind integer modifier return end subroutine subroutine omp_get_schedule(kind, modifier) include 'omp_lib_kinds.h' integer (kind=omp_sched_kind) kind integer modifier kind = omp_sched_static modifier = 0 return end subroutine integer function omp_get_thread_limit() omp_get_thread_limit = 1 return end function subroutine omp_set_max_active_levels( level ) integer level end subroutine integer function omp_get_max_active_levels() omp_get_max_active_levels = 0 return end function integer function omp_get_level() omp_get_level = 0 return end function integer function omp_get_ancestor_thread_num( level ) integer level if ( level .eq. 0 ) then omp_get_ancestor_thread_num = 0 else omp_get_ancestor_thread_num = -1 end if return end function 258 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 integer function omp_get_team_size( level ) integer level if ( level .eq. 0 ) then omp_get_team_size = 1 else omp_get_team_size = -1 end if return end function integer function omp_get_active_level() omp_get_active_level = 0 return end function logical function omp_in_final() omp_in_final = .true. return end function function omp_get_proc_bind() include 'omp_lib_kinds.h' integer (kind=omp_proc_bind_kind) omp_get_proc_bind omp_get_proc_bind = omp_proc_bind_false end function omp_get_proc_bind subroutine omp_set_default_device(device_num) integer device_num return end subroutine integer function omp_get_default_device() omp_get_default_device = 0 return end function integer function omp_get_num_devices() omp_get_num_devices = 0 return end function integer function omp_get_num_teams() omp_get_num_teams = 1 return end function integer function omp_get_team_num() omp_get_team_num = 0 return end function Appendix A Stubs for Runtime Library Routines 259 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 logical function omp_is_initial_device() omp_is_initial_device = .true. return end function subroutine omp_init_lock(lock) ! lock is 0 if the simple lock is not initialized ! -1 if the simple lock is initialized but not set ! 1 if the simple lock is set include 'omp_lib_kinds.h' integer(kind=omp_lock_kind) lock lock = -1 return end subroutine subroutine omp_destroy_lock(lock) include 'omp_lib_kinds.h' integer(kind=omp_lock_kind) lock lock = 0 return end subroutine subroutine omp_set_lock(lock) include 'omp_lib_kinds.h' integer(kind=omp_lock_kind) lock if (lock .eq. -1) then lock = 1 elseif (lock .eq. 1) then print *, 'error: deadlock in using lock variable' stop else print *, 'error: lock not initialized' stop endif return end subroutine 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 260 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 return end subroutine logical function omp_test_lock(lock) include 'omp_lib_kinds.h' integer(kind=omp_lock_kind) lock if (lock .eq. -1) then lock = 1 omp_test_lock = .true. elseif (lock .eq. 1) then omp_test_lock = .false. else print *, 'error: lock not initialized' stop endif return end function subroutine omp_init_nest_lock(nlock) ! nlock is ! 0 if the nestable lock is not initialized ! -1 if the nestable lock is initialized but not set ! 1 if the nestable lock is set ! no use count is maintained include 'omp_lib_kinds.h' integer(kind=omp_nest_lock_kind) nlock nlock = -1 return end subroutine subroutine omp_destroy_nest_lock(nlock) include 'omp_lib_kinds.h' integer(kind=omp_nest_lock_kind) nlock nlock = 0 return end subroutine Appendix A Stubs for Runtime Library Routines 261 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 subroutine omp_set_nest_lock(nlock) include 'omp_lib_kinds.h' integer(kind=omp_nest_lock_kind) nlock if (nlock .eq. -1) nlock = 1 elseif (nlock .eq. print *, 'error: stop else print *, 'error: stop endif then 0) then nested lock not initialized' deadlock using nested lock variable' return end subroutine subroutine omp_unset_nest_lock(nlock) include 'omp_lib_kinds.h' integer(kind=omp_nest_lock_kind) nlock if (nlock .eq. 1) then nlock = -1 elseif (nlock .eq. 0) then print *, 'error: nested lock not initialized' stop else print *, 'error: nested lock not set' stop endif return end subroutine integer function omp_test_nest_lock(nlock) include 'omp_lib_kinds.h' integer(kind=omp_nest_lock_kind) nlock if (nlock .eq. -1) then nlock = 1 omp_test_nest_lock = 1 elseif (nlock .eq. 1) then omp_test_nest_lock = 0 else print *, 'error: nested lock not initialized' stop endif return end function 262 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 double precision function omp_get_wtime() ! this function does not provide a working ! wall clock timer. replace it with a version ! customized for the target machine. omp_get_wtime = 0.0d0 return end function double precision function omp_get_wtick() ! this function does not provide a working ! clock tick function. replace it with ! a version customized for the target machine. double precision one_year parameter (one_year=365.d0*86400.d0) omp_get_wtick = one_year return end function Appendix A Stubs for Runtime Library Routines 263 This page intentionally left blank. 1 264 OpenMP API • Version 4.0 - July 2013 1 APPENDIX B OpenMP C and C++ Grammar 2 3 4 B.1 Notation 5 6 The grammar rules consist of the name for a non-terminal, followed by a colon, followed by replacement alternatives on separate lines. 7 8 The syntactic expression termopt indicates that the term is optional within the replacement. 9 10 The syntactic expression termoptseq is equivalent to term-seqopt with the following additional rules: 11 term-seq : 12 term 13 term-seq term 14 term-seq , term 265 1 B.2 Rules The notation is described in Section 6.1 of the C standard. This grammar appendix shows the extensions to the base language grammar for the OpenMP C and C++ directives. 2 3 4 5 C++ statement-seq: 6 statement 7 8 openmp-directive 9 statement-seq statement statement-seq openmp-directive 10 C++ C90 statement-list: 11 statement 12 13 openmp-directive 14 statement-list statement 15 statement-list openmp-directive C90 C99 block-item: 16 declaration 17 18 statement 19 openmp-directive C99 20 266 OpenMP API • Version 4.0 - July 2013 1 2 statement: /* standard statements */ 3 openmp-construct 4 declaration-definition: 5 6 7 /* Any C or C++ declaration or definition statement */ function-statement: /* C or C++ function definition or declaration */ 8 declarations-definitions-seq: 9 declaration-definition 10 11 12 declarations-definitions-seq declaration-definition openmp-construct: parallel-construct 13 for-construct 14 sections-construct 15 single-construct 16 simd-construct 17 for-simd-construct 18 parallel-for-simd-construct 19 target-data-construct 20 target-construct 21 target-update-construct 22 teams-construct 23 distribute-construct 24 distribute-simd-construct 25 distribute-parallel-for-construct 26 distribute-parallel-for-simd-construct 27 target-teams-construct Appendix B OpenMP C and C++ Grammar 267 1 teams-distribute-construct 2 teams-distribute-simd-construct 3 target-teams-distribute-construct 4 target-teams-distribute-simd-construct 5 teams-distribute-parallel-for-construct 6 target-teams-distribute-parallel-for-construct 7 teams-distribute-parallel-for-simd-construct 8 target-teams-distribute-parallel-for-simd-construct 9 parallel-for-construct 10 parallel-sections-construct 11 task-construct 12 master-construct 13 critical-construct 14 atomic-construct 15 ordered-construct openmp-directive: 16 17 barrier-directive 18 taskwait-directive 19 taskyield-directive 20 flush-directive 21 structured-block: 22 statement parallel-construct: 23 parallel-directive structured-block 24 parallel-directive: 25 # pragma omp parallel parallel-clauseoptseq new-line 26 27 28 268 OpenMP API • Version 4.0 - July 2013 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-clause 10 num_threads ( expression ) 11 copyin ( variable-list ) 12 13 14 15 for-construct: for-directive iteration-statement for-directive: # pragma omp for for-clauseoptseq new-line 16 17 18 for-clause: unique-for-clause 19 data-privatization-clause 20 data-privatization-in-clause 21 data-privatization-out-clause 22 data-reduction-clause 23 nowait 24 unique-for-clause: 25 ordered 26 schedule ( schedule-kind ) 27 schedule ( schedule-kind , expression ) 28 collapse ( expression ) Appendix B OpenMP C and C++ Grammar 269 schedule-kind: 1 2 static 3 dynamic 4 guided 5 auto 6 runtime sections-construct: 7 sections-directive section-scope 8 sections-directive: 9 # pragma omp sections sections-clauseoptseq new-line 10 sections-clause: 11 12 data-privatization-clause 13 data-privatization-in-clause 14 data-privatization-out-clause 15 data-reduction-clause 16 nowait 17 section-scope: { section-sequence } 18 section-sequence: 19 section-directiveopt structured-block 20 section-sequence section-directive structured-block 21 section-directive: 22 # pragma omp section new-line 23 single-construct: 24 single-directive structured-block 25 single-directive: 26 # pragma omp single single-clauseoptseq new-line 27 28 270 OpenMP API • Version 4.0 - July 2013 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 ) simd-construct: simd-directive iteration-statement simd-directive: # pragma omp simd simd-clauseoptseq new-line simd-clause: 13 collapse ( expression ) 14 aligned-clause 15 linear-clause 16 uniform-clause 17 data-reduction-clause 18 inbranch-clause 19 20 inbranch-clause: 21 inbranch 22 notinbranch 23 24 25 uniform-clause: uniform ( variable-list ) linear-clause: 26 linear ( variable-list ) 27 linear ( variable-list : expression ) Appendix B OpenMP C and C++ Grammar 271 aligned-clause: 1 2 aligned ( variable-list ) 3 aligned ( variable-list : expression ) declare-simd-construct: 4 declare-simd-directive-seq function-statement 5 6 declare-simd-directive-seq: 7 declare-simd-directive 8 declare-simd-directive-seq declare-simd-directive declare-simd-directive: 9 # pragma omp declare simd declare-simd-clauseoptseq new-line 10 declare-simd-clause: 11 12 simdlen ( expression ) 13 aligned-clause 14 linear-clause 15 uniform-clause 16 data-reduction-clause 17 inbranch-clause for-simd-construct: 18 for-simd-directive iteration-statement 19 20 for-simd-directive: 21 # pragma omp for simd for-simd-clauseoptseq new-line 22 23 for-simd-clause: 24 for-clause 25 simd-clause parallel-for-simd-construct: 26 parallel-for-simd-directive iteration-statement 27 272 OpenMP API • Version 4.0 - July 2013 1 2 parallel-for-simd-directive: # pragma omp parallel for simd parallel-for-simd-clauseoptseq new-line 3 parallel-for-simd-clause: 4 parallel-for-clause 5 simd-clause 6 7 8 9 target-data-construct: target-data-directive structured-block target-data-directive: # pragma omp target data target-data-clauseoptseq new-line 10 target-data-clause: 11 device-clause 12 map-clause 13 if-clause 14 device-clause: 15 16 17 18 device ( expression ) map-clause: map ( map-typeopt variable-array-section-list ) map-type: 19 alloc: 20 to: 21 from: 22 tofrom: 23 target-construct: 24 25 26 target-directive structured-block target-directive: # pragma omp target target-clauseoptseq new-line 27 Appendix B OpenMP C and C++ Grammar 273 target-clause: 1 2 device-clause 3 map-clause 4 if-clause target-update-construct: 5 target-update-directive structured-block 6 target-update-directive: 7 # pragma omp target update target-update-clauseseq new-line 8 9 target-update-clause: 10 motion-clause 11 device-clause 12 if-clause 13 motion-clause: 14 to ( variable-array-section-list ) 15 from ( variable-array-section-list ) declare-target-construct: 16 declare-target-directive declarations-definitions-seq end-declare-target-directive 17 declare-target-directive: 18 # pragma omp declare target new-line 19 end-declare-target-directive: 20 # pragma omp end declare target new-line 21 teams-construct: 22 teams-directive structured-block 23 teams-directive: 24 # pragma omp teams teams-clauseoptseq new-line 25 26 27 274 OpenMP API • Version 4.0 - July 2013 1 teams-clause: 2 num_teams ( expression ) 3 thread_limit ( expression ) 4 data-default-clause 5 data-privatization-clause 6 data-privatization-in-clause 7 data-sharing-clause 8 data-reduction-clause 9 10 11 12 13 distribute-construct: distribute-directive iteration-statement distribute-directive: # pragma omp distribute distribute-clauseoptseq new-line distribute-clause: 14 data-privatization-clause 15 data-privatization-in-clause 16 collapse ( expression ) 17 dist_schedule ( static ) 18 dist_schedule ( static , expression ) 19 20 21 22 distribute-simd-construct: distribute-simd-directive iteration-statement distribute-simd-directive: #pragma omp distribute simd distribute-simd-clauseoptseq new-line 23 distribute-simd-clause: 24 distribute-clause 25 simd-clause 26 27 distribute-parallel-for-construct: distribute-parallel-for-directive iteration-statement 28 Appendix B OpenMP C and C++ Grammar 275 1 distribute-parallel-for-directive: 2 3 #pragma omp distribute parallel for distribute-parallel-for-clauseoptseq new-line distribute-parallel-for-clause: 4 5 distribute-clause 6 parallel-for-clause distribute-parallel-for-simd-construct: 7 distribute-parallel-for-simd-directive iteration-statement 8 distribute-parallel-for-simd-directive: 9 #pragma omp distribute parallel for distribute-parallel-for-simdclauseoptseq new-line 10 11 distribute-parallel-for-simd-clause: 12 13 distribute-clause 14 parallel-for-simd-clause target-teams-construct: 15 target-teams-directive iteration-statement 16 target-teams-directive: 17 #pragma omp target teams target-teams-clauseoptseq new-line 18 target-teams-clause: 19 20 target-clause 21 teams-clause teams-distribute-construct: 22 teams-distribute-directive iteration-statement 23 teams-distribute-directive: 24 #pragma omp teams distribute teams-distribute-clauseoptseq new-line 25 teams-distribute-clause: 26 27 teams-clause 28 distribute-clause teams-distribute-simd-construct: 29 teams-distribute-simd-directive iteration-statement 30 31 276 OpenMP API • Version 4.0 - July 2013 1 2 3 4 teams-distribute-simd-directive: #pragma omp teams distribute simd teams-distribute-simd-clauseoptseq new-line teams-distribute-simd-clause: 5 teams-clause 6 distribute-simd-clause 7 8 9 10 11 12 target-teams-distribute-construct: target-teams-distribute-directive iteration-statement target-teams-distribute-directive: #pragma omp target teams distribute target-teams-distribute-clauseoptseq new-line target-teams-distribute-clause: 13 target-clause 14 teams-distribute-clause 15 16 17 18 19 20 target-teams-distribute-simd-construct: target-teams-distribute-simd-directive iteration-statement target-teams-distribute-simd-directive: #pragma omp target teams distribute simd target-teams-distributesimd-clauseoptseq new-line target-teams-distribute-simd-clause: 21 target-clause 22 teams-distribute-simd-clause 23 24 25 26 27 28 teams-distribute-parallel-for-construct: teams-distribute-parallel-for-directive iteration-statement teams-distribute-parallel-for-directive: #pragma omp teams distribute parallel for teams-distributeparallel-for-clauseoptseq new-line teams-distribute-parallel-for-clause: 29 teams-clause 30 distribute-parallel-for-clause 31 Appendix B OpenMP C and C++ Grammar 277 target-teams-distribute-parallel-for-construct: 1 target-teams-distribute-parallel-for-directive iteration-statement 2 target-teams-distribute-parallel-for-directive: 3 #pragma omp teams distribute parallel for target-teams-distributeparallel-for-clauseoptseq new-line 4 5 target-teams-distribute-parallel-for-clause: 6 7 target-clause 8 teams-distribute-parallel-for-clause teams-distribute-parallel-for-simd-construct: 9 teams-distribute-parallel-for-simd-directive iteration-statement 10 teams-distribute-parallel-for-simd-directive: 11 #pragma omp teams distribute parallel for simd teams-distributeparallel-for-simd-clauseoptseq new-line 12 13 teams-distribute-parallel-for-simd-clause: 14 15 teams-clause 16 distribute-parallel-for-simd-clause target-teams-distribute-parallel-for-simd-construct: 17 target-teams-distribute-parallel-for-simd-directive iteration-statement 18 target-teams-distribute-parallel-for-simd-directive: 19 #pragma omp target teams distribute parallel for simd targetteams-distribute-parallel-for-simd-clauseoptseq new-line 20 21 target-teams-distribute-parallel-for-simd-clause: 22 23 target-clause 24 teams-distribute-parallel-for-simd-clause task-construct: 25 task-directive structured-block 26 task-directive: 27 # pragma omp task task-clauseoptseq new-line 28 29 30 31 278 OpenMP API • Version 4.0 - July 2013 1 task-clause: 2 unique-task-clause 3 data-default-clause 4 data-privatization-clause 5 data-privatization-in-clause 6 data-sharing-clause 7 unique-task-clause: 8 if-clause 9 final( scalar-expression ) 10 untied 11 mergeable 12 depend ( dependence-type : variable-array-section-list ) 13 dependence-type: 14 in 15 out 16 inout 17 18 19 20 21 22 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 23 unique-for-clause 24 data-default-clause 25 data-privatization-clause 26 data-privatization-in-clause 27 data-privatization-out-clause Appendix B OpenMP C and C++ Grammar 279 1 data-sharing-clause 2 data-reduction-clause 3 parallel-sections-construct: parallel-sections-directive section-scope 4 parallel-sections-directive: 5 # pragma omp parallel sections parallel-sections-clauseoptseq new-line 6 parallel-sections-clause: 7 8 unique-parallel-clause 9 data-default-clause 10 data-privatization-clause 11 data-privatization-in-clause 12 data-privatization-out-clause 13 data-sharing-clause 14 data-reduction-clause master-construct: 15 master-directive structured-block 16 master-directive: 17 # pragma omp master new-line 18 critical-construct: 19 critical-directive structured-block 20 critical-directive: 21 # pragma omp critical region-phraseopt new-line 22 region-phrase: 23 ( identifier ) 24 barrier-directive: 25 # pragma omp barrier new-line 26 taskwait-directive: 27 # pragma omp taskwait new-line 28 280 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 taskgroup-construct: taskgroup-directive structured-block taskgroup-directive: # pragma omp taskgroup new-line taskyield-directive: # pragma omp taskyield new-line atomic-construct: atomic-directive expression-statement atomic-directive structured block atomic-directive: # pragma omp atomic atomic-clauseopt seq_cst-clauseopt new-line atomic-clause: 13 read 14 write 15 update 16 capture 17 seq-cst-clause: 18 seq_cst 19 flush-directive: 20 21 22 23 24 25 26 27 28 # pragma omp flush flush-varsopt new-line flush-vars: ( variable-list ) ordered-construct: ordered-directive structured-block ordered-directive: # pragma omp ordered new-line cancel-directive: # pragma omp cancel construct-type-clause if-clauseopt new-line Appendix B OpenMP C and C++ Grammar 281 construct-type-clause: 1 2 parallel 3 sections 4 for 5 taskgroup cancellation-point-directive: 6 # pragma omp cancellation point construct-type-clause new-line 7 declaration: 8 /* standard declarations */ 9 10 threadprivate-directive 11 declare-simd-directive 12 declare-target-construct 13 declare-reduction-directive threadprivate-directive: 14 # pragma omp threadprivate ( variable-list ) new-line 15 declare-reduction-directive: 16 17 18 # pragma omp declare reduction (reduction-identifier : reduction-type-list : expression ) initializer-clauseopt new-line reduction-identifier: 19 C identifier 20 C C++ id-expression 21 C++ C/C++ one of: + * - & ^ | && || min max 22 C/C++ 282 OpenMP API • Version 4.0 - July 2013 1 reduction-type-list: 2 type-id 3 reduction-type-list, type-id 4 initializer-clause: C 5 initializer ( identifier = initializer ) 6 initializer ( identifier ( argument-expression-list ) ) C C++ 7 initializer ( identifier initializer ) 8 initializer ( id-expression ( expression-list ) ) C++ 9 10 data-default-clause: 11 default ( shared ) 12 default ( none ) 13 14 15 16 17 18 19 20 21 22 23 24 data-privatization-clause: private ( variable-list ) data-privatization-in-clause: firstprivate ( variable-list ) data-privatization-out-clause: lastprivate ( variable-list ) data-sharing-clause: shared ( variable-list ) data-reduction-clause: reduction ( reduction-identifier : variable-list ) if-clause: if ( scalar-expression ) Appendix B OpenMP C and C++ Grammar 283 C array-section: 1 identifier array-section-subscript 2 3 variable-list: 4 identifier variable-list , identifier 5 variable-array-section-list: 6 7 identifier 8 array-section 9 variable-array-section-list , identifier variable-array-section-list , array-section 10 C C++ array-section: 11 id-expression array-section-subscript 12 variable-list: 13 id-expression 14 variable-list , id-expression 15 variable-array-section-list: 16 17 id-expression 18 array-section 19 variable-array-section-list , id-expression 20 variable-array-section-list , array-section C++ 21 22 23 24 284 OpenMP API • Version 4.0 - July 2013 1 array-section-subscript: 2 array-section-subscript [ expressionopt : expressionopt ] 3 array-section-subscript [ expression ] 4 [ expressionopt : expressionopt ] 5 [ expression ] Appendix B OpenMP C and C++ Grammar 285 This page intentionally left blank. 1 286 OpenMP API • Version 4.0 - July 2013 1 APPENDIX C 2 Interface Declarations 3 4 5 6 This appendix gives examples of the C/C++ header file, the Fortran include file and Fortran module that shall be provided by implementations as specified in Chapter 3. It also includes an example of a Fortran 90 generic interface for a library routine. This is a non-normative section, implementation files may differ. 287 1 C.1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 Example of the omp.h Header File #ifndef _OMP_H_DEF #define _OMP_H_DEF /* * define the lock data types */ typedef void *omp_lock_t; typedef void *omp_nest_lock_t; /* * define the schedule kinds */ typedef enum omp_sched_t { omp_sched_static = 1, omp_sched_dynamic = 2, omp_sched_guided = 3, omp_sched_auto = 4 /* , Add vendor specific schedule constants here */ } omp_sched_t; /* * define the proc bind values */ typedef enum omp_proc_bind_t { omp_proc_bind_false = 0, omp_proc_bind_true = 1, omp_proc_bind_master = 2, omp_proc_bind_close = 3, omp_proc_bind_spread = 4 } omp_proc_bind_t; /* * exported OpenMP functions */ #ifdef __cplusplus extern "C" { #endif extern extern extern extern extern extern extern 288 void int int int int int void 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); OpenMP API • Version 4.0 - July 2013 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 extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern extern int omp_get_dynamic(void); void omp_set_nested(int nested); int omp_get_cancellation(void); int omp_get_nested(void); void omp_set_schedule(omp_sched_t kind, int modifier); void omp_get_schedule(omp_sched_t *kind, int *modifier); int omp_get_thread_limit(void); void omp_set_max_active_levels(int max_active_levels); int omp_get_max_active_levels(void); int omp_get_level(void); int omp_get_ancestor_thread_num(int level); int omp_get_team_size(int level); int omp_get_active_level(void); int omp_in_final(void); omp_proc_bind_t omp_get_proc_bind(void); void omp_set_default_device(int device_num); int omp_get_default_device(void); int omp_get_num_devices(void); int omp_get_num_teams(void); int omp_get_team_num(void); int omp_is_initial_device(void); extern extern extern extern extern void void void void int omp_init_lock(omp_lock_t *lock); omp_destroy_lock(omp_lock_t *lock); omp_set_lock(omp_lock_t *lock); omp_unset_lock(omp_lock_t *lock); omp_test_lock(omp_lock_t *lock); extern extern extern extern extern void void void void int omp_init_nest_lock(omp_nest_lock_t *lock); omp_destroy_nest_lock(omp_nest_lock_t *lock); omp_set_nest_lock(omp_nest_lock_t *lock); omp_unset_nest_lock(omp_nest_lock_t *lock); omp_test_nest_lock(omp_nest_lock_t *lock); extern double omp_get_wtime(void); extern double omp_get_wtick(void); #ifdef __cplusplus } #endif #endif Appendix C Interface Declarations 289 1 C.2 2 Example of an Interface Declaration include File omp_lib_kinds.h: 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 integer omp_lock_kind integer omp_nest_lock_kind ! this selects an integer that is large enough to hold a 64 bit integer parameter ( omp_lock_kind = selected_int_kind( 10 ) ) parameter ( omp_nest_lock_kind = selected_int_kind( 10 ) ) integer omp_sched_kind ! this selects an integer that is large enough to hold a 32 bit integer parameter ( omp_sched_kind = selected_int_kind( 8 ) ) integer ( omp_sched_kind ) omp_sched_static parameter ( omp_sched_static = 1 ) integer ( omp_sched_kind ) omp_sched_dynamic parameter ( omp_sched_dynamic = 2 ) integer ( omp_sched_kind ) omp_sched_guided parameter ( omp_sched_guided = 3 ) integer ( omp_sched_kind ) omp_sched_auto parameter ( omp_sched_auto = 4 ) integer omp_proc_bind_kind parameter ( omp_proc_bind_kind = selected_int_kind( 8 ) ) integer ( omp_proc_bind_kind ) omp_proc_bind_false parameter ( omp_proc_bind_false = 0 ) integer ( omp_proc_bind_kind ) omp_proc_bind_true parameter ( omp_proc_bind_true = 1 ) integer ( omp_proc_bind_kind ) omp_proc_bind_master parameter ( omp_proc_bind_master = 2 ) integer ( omp_proc_bind_kind ) omp_proc_bind_close parameter ( omp_proc_bind_close = 3 ) integer ( omp_proc_bind_kind ) omp_proc_bind_spread parameter ( omp_proc_bind_spread = 4 ) omp_lib.h: 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 ! default integer type assumed below ! default logical type assumed below ! OpenMP API v4.0 include 'omp_lib_kinds.h' integer openmp_version parameter ( openmp_version = 201307 ) external external integer external integer external integer external 290 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 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 integer omp_get_num_procs external omp_in_parallel logical omp_in_parallel external omp_set_dynamic external omp_get_dynamic logical omp_get_dynamic external omp_get_cancellation integer omp_get_cancellation external omp_set_nested external omp_get_nested logical omp_get_nested external omp_set_schedule external omp_get_schedule external omp_get_thread_limit integer omp_get_thread_limit external omp_set_max_active_levels external omp_get_max_active_levels integer omp_get_max_active_levels external omp_get_level integer omp_get_level external omp_get_ancestor_thread_num integer omp_get_ancestor_thread_num external omp_get_team_size integer omp_get_team_size external omp_get_active_level integer omp_get_active_level external omp_set_default_device external omp_get_default_device integer omp_get_default_device external omp_get_num_devices integer omp_get_num_devices external omp_get_num_teams integer omp_get_num_teams external omp_get_team_num integer omp_get_team_num external omp_is_initial_device logical omp_is_initial_device external omp_in_final logical omp_in_final integer ( omp_proc_bind_kind ) omp_get_proc_bind external omp_get_proc_bind external external external external external logical omp_init_lock omp_destroy_lock omp_set_lock omp_unset_lock omp_test_lock omp_test_lock external omp_init_nest_lock external omp_destroy_nest_lock external omp_set_nest_lock Appendix C Interface Declarations 291 1 2 3 4 5 6 7 8 external omp_unset_nest_lock external omp_test_nest_lock integer omp_test_nest_lock external omp_get_wtick double precision omp_get_wtick external omp_get_wtime double precision omp_get_wtime 9 292 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 C.3 Example of a Fortran Interface Declaration module ! the "!" of this comment starts in column 1 !23456 module omp_lib_kinds integer, parameter :: omp_lock_kind = selected_int_kind( 10 ) integer, parameter :: omp_nest_lock_kind = selected_int_kind( 10 ) integer, parameter :: omp_sched_kind = selected_int_kind( 8 ) integer(kind=omp_sched_kind), parameter :: & omp_sched_static = 1 integer(kind=omp_sched_kind), parameter :: & omp_sched_dynamic = 2 integer(kind=omp_sched_kind), parameter :: & omp_sched_guided = 3 integer(kind=omp_sched_kind), parameter :: & omp_sched_auto = 4 integer, parameter :: omp_proc_bind_kind = selected_int_kind( 8 ) integer (kind=omp_proc_bind_kind), parameter :: & omp_proc_bind_false = 0 integer (kind=omp_proc_bind_kind), parameter :: & omp_proc_bind_true = 1 integer (kind=omp_proc_bind_kind), parameter :: & omp_proc_bind_master = 2 integer (kind=omp_proc_bind_kind), parameter :: & omp_proc_bind_close = 3 integer (kind=omp_proc_bind_kind), parameter :: & omp_proc_bind_spread = 4 end module omp_lib_kinds module omp_lib use omp_lib_kinds ! OpenMP API v4.0 integer, parameter :: openmp_version = 201307 interface subroutine omp_set_num_threads (number_of_threads_expr) integer, intent(in) :: number_of_threads_expr end subroutine omp_set_num_threads function omp_get_num_threads () integer :: omp_get_num_threads end function omp_get_num_threads function omp_get_max_threads () integer :: omp_get_max_threads end function omp_get_max_threads Appendix C Interface Declarations 293 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 function omp_get_thread_num () integer :: omp_get_thread_num end function omp_get_thread_num function omp_get_num_procs () integer :: omp_get_num_procs end function omp_get_num_procs function omp_in_parallel () logical :: omp_in_parallel end function omp_in_parallel subroutine omp_set_dynamic (enable_expr) logical, intent(in) ::enable_expr end subroutine omp_set_dynamic function omp_get_dynamic () logical :: omp_get_dynamic end function omp_get_dynamic function omp_get_cancellation () integer :: omp_get_cancellation end function omp_get_cancellation subroutine omp_set_nested (enable_expr) logical, intent(in) :: enable_expr end subroutine omp_set_nested function omp_get_nested () logical :: omp_get_nested end function omp_get_nested subroutine omp_set_schedule (kind, modifier) use omp_lib_kinds integer(kind=omp_sched_kind), intent(in) :: kind integer, intent(in) :: modifier end subroutine omp_set_schedule subroutine omp_get_schedule (kind, modifier) use omp_lib_kinds integer(kind=omp_sched_kind), intent(out) :: kind integer, intent(out)::modifier end subroutine omp_get_schedule function omp_get_thread_limit() integer :: omp_get_thread_limit end function omp_get_thread_limit subroutine omp_set_max_active_levels(var) integer, intent(in) :: var end subroutine omp_set_max_active_levels 294 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 function omp_get_max_active_levels() integer :: omp_get_max_active_levels end function omp_get_max_active_levels function omp_get_level() integer :: omp_get_level end function omp_get_level function omp_get_ancestor_thread_num(level) integer, intent(in) :: level integer :: omp_get_ancestor_thread_num end function omp_get_ancestor_thread_num function omp_get_team_size(level) integer, intent(in) :: level integer :: omp_get_team_size end function omp_get_team_size function omp_get_active_level() integer :: omp_get_active_level end function omp_get_active_level function omp_in_final() logical omp_in_final end function omp_in_final function omp_get_proc_bind( ) include 'omp_lib_kinds.h' integer (kind=omp_proc_bind_kind) omp_get_proc_bind omp_get_proc_bind = omp_proc_bind_false end function omp_get_proc_bind subroutine omp_set_default_device (device_num) integer :: device_num end subroutine omp_set_default_device function omp_get_default_device () integer :: omp_get_default_device end function omp_get_default_device function omp_get_num_devices () integer :: omp_get_num_devices end function omp_get_num_devices function omp_get_num_teams () integer :: omp_get_num_teams end function omp_get_num_teams function omp_get_team_num () integer :: omp_get_team_num end function omp_get_team_num Appendix C Interface Declarations 295 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 function omp_is_initial_device () logical :: omp_is_initial_device end function omp_is_initial_device subroutine omp_init_lock (var) use omp_lib_kinds integer (kind=omp_lock_kind), intent(out) :: var end subroutine omp_init_lock subroutine omp_destroy_lock (var) use omp_lib_kinds integer (kind=omp_lock_kind), intent(inout) :: var end subroutine omp_destroy_lock subroutine omp_set_lock (var) use omp_lib_kinds integer (kind=omp_lock_kind), intent(inout) :: var end subroutine omp_set_lock subroutine omp_unset_lock (var) use omp_lib_kinds integer (kind=omp_lock_kind), intent(inout) :: var end subroutine omp_unset_lock function omp_test_lock (var) use omp_lib_kinds logical :: omp_test_lock integer (kind=omp_lock_kind), intent(inout) :: var end function omp_test_lock subroutine omp_init_nest_lock (var) use omp_lib_kinds integer (kind=omp_nest_lock_kind), intent(out) :: var end subroutine omp_init_nest_lock subroutine omp_destroy_nest_lock (var) use omp_lib_kinds integer (kind=omp_nest_lock_kind), intent(inout) :: var end subroutine omp_destroy_nest_lock subroutine omp_set_nest_lock (var) use omp_lib_kinds integer (kind=omp_nest_lock_kind), intent(inout) :: var end subroutine omp_set_nest_lock subroutine omp_unset_nest_lock (var) use omp_lib_kinds integer (kind=omp_nest_lock_kind), intent(inout) :: var end subroutine omp_unset_nest_lock function omp_test_nest_lock (var) use omp_lib_kinds integer :: omp_test_nest_lock integer (kind=omp_nest_lock_kind), intent(inout) :: var 296 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 end function omp_test_nest_lock function omp_get_wtick () double precision :: omp_get_wtick end function omp_get_wtick function omp_get_wtime () double precision :: omp_get_wtime end function omp_get_wtime end interface end module omp_lib Appendix C Interface Declarations 297 C.4 2 Example of a Generic Interface for a Library Routine 3 4 Any of the OpenMP runtime library routines that take an argument may be extended with a generic interface so arguments of different KIND type can be accommodated. 5 6 The OMP_SET_NUM_THREADS interface could be specified in the omp_lib module as follows: 1 interface omp_set_num_threads subroutine omp_set_num_threads_4(number_of_threads_expr) use omp_lib_kinds integer(4), intent(in) :: number_of_threads_expr end subroutine omp_set_num_threads_4 subroutine omp_set_num_threads_8(number_of_threads_expr) use omp_lib_kinds integer(8), intent(in) :: number_of_threads_expr end subroutine omp_set_num_threads_8 end interface omp_set_num_threads 7 8 298 OpenMP API • Version 4.0 - July 2013 1 APPENDIX D 3 OpenMP ImplementationDefined Behaviors 4 5 6 7 This appendix summarizes the behaviors that are described as implementation defined in this API. Each behavior is cross-referenced back to its description in the main specification. An implementation is required to define and document its behavior in these cases. 8 9 • Processor: a hardware unit that is implementation defined (see Section 1.2.1 on page 2 2). 10 11 • Device: an implementation defined logical execution engine (see Section 1.2.1 on 12 13 14 15 • Memory model: the minimum size at which a memory update may also read and 16 17 18 • Memory model: Implementations are allowed to relax the ordering imposed by 19 20 21 22 • Internal control variables: the initial values of dyn-var, nthreads-var, run-sched-var, 23 24 25 26 • Dynamic adjustment of threads: providing the ability to dynamically adjust the 27 28 29 30 • Thread affinity: With T<=P, when T does not divide P evenly, the assignment of the page 2). write back adjacent variables that are part of another variable (as array or structure elements) is implementation defined but is no larger than required by the base language (see Section 1.4.1 on page 17). implicit flush operations when the result is only visible to programs using nonsequentially consistent atomic directives (see Section 1.4.4 on page 20). def-sched-var, bind-var, stacksize-var, wait-policy-var, thread-limit-var, max-activelevels-var, place-partition-var, and default-device-var are implementation defined (see Section 2.3.2 on page 36). number of threads is implementation defined . Implementations are allowed to deliver fewer threads (but at least one) than indicated in Algorithm 2-1 even if dynamic adjustment is disabled (see Section 2.5.1 on page 47). remaining P-T*S places into subpartitions is implementation defined. With T>P, when P does not divide T evenly, the assignment of the remaining T-P*S threads into places is implementation defined. The determination of whether the affinity request 299 1 2 3 can be fulfilled is implementation defined. If not, the number of threads in the team and their mapping to places become implementation defined (see Section 2.5.2 on page 49). 4 5 6 7 • Loop directive: the integer type (or kind, for Fortran) used to compute the iteration 8 9 • sections construct: the method of scheduling the structured blocks among threads 10 11 • single construct: the method of choosing a thread to execute the structured block 12 13 14 15 16 • simd construct: the integer type (or kind, for Fortran) used to compute the iteration 17 18 19 20 • declare simd construct: if the simdlen clause is not specified, the number of 21 22 23 24 25 • teams construct: the number of teams that are created is implementation defined but 26 27 • If no dist_schedule clause is specified then the schedule for the distribute 28 29 30 31 32 33 • atomic construct: a compliant implementation may enforce exclusive access 34 35 • omp_set_num_threads routine: if the argument is not a positive integer the 36 37 38 • omp_set_schedule routine: for implementation specific schedule types, the count of a 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 2.7.1 on page 53. in the team is implementation defined (see Section 2.7.2 on page 60). is implementation defined (see Section 2.7.3 on page 63) count for the collapsed loop is implementation defined. The number of iterations that are executed concurrently at any given time is implementation defined. If the aligned clause is not specified, the assumed alignment is implementation defined (see Section 2.8.1 on page 68). concurrent arguments for the function is implementation defined. If the aligned clause is not specified, the assumed alignment is implementation defined (see Section 2.8.2 on page 72). less than or equal to the value of the num_teams clause if specified. The maximum number of threads participating in the contention group that each team initiates is implementation defined but less than or equal to the value of the thread_limit clause if specified (see Section 2.9.5 on page 86). construct is implementation defined (see Section 2.9.6 on page 88). between atomic regions that update different storage locations. The circumstances under which this occurs are implementation defined. If the storage location designated by x is not size-aligned (that is, if the byte alignment of x is not a multiple of the size of x), then the behavior of the atomic region is implementation defined (see Section 2.12.6 on page 127). behavior is implementation defined (see Section 3.2.1 on page 189). values and associated meanings of the second argument are implementation defined. (see Section 3.2.12 on page 203). 300 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 • omp_set_max_active_levels routine: when called from within any explicit 6 7 8 9 • omp_get_max_active_levels routine: 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 and the behavior is implementation defined. If the argument is not a non-negative integer then the behavior is implementation defined (see Section 3.2.15 on page 207). parallel region the binding thread set (and binding region, if required) for the omp_get_max_active_levels region is implementation defined (see Section 3.2.16 on page 209). 10 11 12 • OMP_SCHEDULE environment variable: if the value of the variable does not 13 14 15 16 • OMP_NUM_THREADS environment variable: if any value of the list specified in the 17 18 19 20 21 • OMP_PROC_BIND environment variable: if the value is not true, false, or a 22 23 • OMP_DYNAMIC environment variable: if the value is neither true nor false the 24 25 • OMP_NESTED environment variable: if the value is neither true nor false the 26 27 28 • OMP_STACKSIZE environment variable: if the value does not conform to the 29 30 • OMP_WAIT_POLICY environment variable: the details of the ACTIVE and 31 32 33 34 • OMP_MAX_ACTIVE_LEVELS environment variable: if the value is not a non- 35 36 37 38 • OMP_THREAD_LIMIT environment variable: if the requested value is greater than 39 40 41 • OMP_PLACES environment variable: the meaning of the numbers specified in the conform to the specified format then the result is implementation defined (see Section 4.1 on page 238). OMP_NUM_THREADS environment variable leads to a number of threads that is greater than the implementation can support, or if any value is not a positive integer, then the result is implementation defined (see Section 4.2 on page 239). comma separated list of master, close, or spread, the behavior is implementation defined. The behavior is also implementation defined if an initial thread cannot be bound to the first place in the OpenMP place list (see Section 4.4 on page 241). behavior is implementation defined (see Section 4.3 on page 240). behavior is implementation defined (see Section on page 241). specified format or the implementation cannot provide a stack of the specified size then the behavior is implementation defined (see Section 4.7 on page 244). PASSIVE behaviors are implementation defined (see Section 4.8 on page 245). negative integer or is greater than the number of parallel levels an implementation can support then the behavior is implementation defined (see Section 4.9 on page 245). 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 4.10 on page 246). environment variable and how the numbering is done are implementation defined. The precise definitions of the abstract names are implementation defined. An Appendix D OpenMP Implementation-Defined Behaviors 301 1 2 3 4 5 6 7 8 9 10 11 implementation may add implementation-defined abstract names as appropriate for the target platform. When creating a place list of n elements by appending the number n to an abstract name, the determination of which resources to include in the place list is implementation defined. When requesting more resources than available, the length of the place list is also implementation defined. The behavior of the program is implementation defined when the execution environment cannot map a numerical value (either explicitly defined or implicitly derived from an interval) within the OMP_PLACES list to a processor on the target platform, or if it maps to an unavailable processor. The behavior is also implementation defined when the OMP_PLACES environment variable is defined using an abstract name (see Section 4.5 on page 241). 12 13 14 • Thread affinity policy: if the affinity request for a parallel construct cannot be fulfilled, the behavior of the thread affinity policy is implementation defined for that parallel construct. Fortran 15 16 17 18 • threadprivate directive: if the conditions for values of data in the threadprivate 19 20 21 22 23 • shared clause: passing a shared variable to a non-intrinsic procedure may result in 24 25 26 27 28 • Runtime library definitions: it is implementation defined whether the include file objects of threads (other than an initial thread) to persist between two consecutive active parallel regions do not all hold, the allocation status of an allocatable variable in the second region is implementation defined (see Section 2.14.2 on page 150). 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 2.14.3.2 on page 157). 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 188). Fortran 302 OpenMP API • Version 4.0 - July 2013 1 APPENDIX E 2 Features History 3 4 This appendix summarizes the major changes between recent versions of the OpenMP API since version 2.5. 5 E.1 Version 3.1 to 4.0 Differences 6 7 • Various changes throughout the specification were made to provide initial support of 8 9 • C/C++ array syntax was extended to support array sections (see Section 2.4 on page Fortran 2003 (see Section 1.6 on page 22). 42). 10 11 12 13 • The proc_bind clause (see Section 2.5.2 on page 49), the OMP_PLACES 14 15 • SIMD constructs were added to support SIMD parallelism (see Section 2.8 on page 16 17 18 19 20 • Device constructs (see Section 2.9 on page 77), the OMP_DEFAULT_DEVICE 21 22 • Implementation defined task scheduling points for untied tasks were removed (see 23 24 • The depend clause (see Section 2.11.1.1 on page 116) was added to support task 25 26 • The taskgroup construct (see Section 2.12.5 on page 126) was added to support environment variable (see Section 4.5 on page 241), and the omp_get_proc_bind runtime routine (see Section 3.2.22 on page 216) were added to support thread affinity policies. 68). environment variable (see Section 4.13 on page 248), the omp_set_default_device, omp_get_default_device, omp_get_num_devices, omp_get_num_teams, omp_get_team_num, and omp_is_initial_device routines were added to support execution on devices. Section 2.11.3 on page 118). dependences. more flexible deep task synchronization. 303 1 2 3 • The reduction clause (see Section 2.14.3.6 on page 167) was extended and the 4 5 6 7 • The atomic construct (see Section 2.12.6 on page 127) was extended to support declare reduction construct (see Section 2.15 on page 180) was added to support user defined reductions. atomic swap with the capture clause, to allow new atomic update and capture forms, and to support sequentially consistent atomic operations with a new seq_cst clause. 8 9 10 11 12 • The cancel construct (see Section 2.13.1 on page 140), the cancellation 13 14 15 • The OMP_DISPLAY_ENV environment variable (see Section 4.12 on page 247) was 16 • Examples (previously Appendix A) were moved to a separate document. 17 point construct (see Section 2.13.2 on page 143), the omp_get_cancellation runtime routine (see Section 3.2.9 on page 199) and the OMP_CANCELLATION environment variable (see Section 4.11 on page 246) were added to support the concept of cancellation. added to display the value of ICVs associated with the OpenMP environment variables. E.2 Version 3.0 to 3.1 Differences 18 19 • The final and mergeable clauses (see Section 2.11.1 on page 113) were added to 20 21 • The taskyield construct (see Section 2.11.2 on page 117) was added to allow 22 23 24 • The atomic construct (see Section 2.12.6 on page 127) was extended to include 25 26 • Data environment restrictions were changed to allow intent(in) and const- 27 28 29 • Data environment restrictions were changed to allow Fortran pointers in 30 • New reduction operators min and max were added for C and C++ 31 32 33 34 • The nesting restrictions in Section 2.16 on page 186 were clarified to disallow 35 36 • The omp_in_final runtime library routine (see Section 3.2.21 on page 215) was the task construct to support optimization of task data environments. user-defined task scheduling points. read, write, and capture forms, and an update clause was added to apply the already existing form of the atomic construct. qualified types for the firstprivate clause (see Section 2.14.3.4 on page 162). firstprivate (see Section 2.14.3.4 on page 162) and lastprivate (see Section 2.14.3.5 on page 164). closely-nested OpenMP regions within an atomic region. This allows an atomic region to be consistently defined with other OpenMP regions so that they include all the code in the atomic construct. added to support specialization of final task regions. 304 OpenMP API • Version 4.0 - July 2013 1 2 3 4 5 • The nthreads-var ICV has been modified to be a list of the number of threads to use 6 7 8 • The bind-var ICV has been added, which controls whether or not threads are bound 9 • Descriptions of examples (see Appendix A on page 221) were expanded and clarified. at each nested parallel region level. The value of this ICV is still set with the OMP_NUM_THREADS environment variable (see Section 4.2 on page 239), but the algorithm for determining the number of threads used in a parallel region has been modified to handle a list (see Section 2.5.1 on page 47). to processors (see Section 2.3.1 on page 35). The value of this ICV can be set with the OMP_PROC_BIND environment variable (see Section 4.4 on page 241). • Replaced incorrect use of omp_integer_kind in Fortran interfaces (see 10 11 12 13 Section C.3 on page 293 and Section C.4 on page 298) with selected_int_kind(8). E.3 Version 2.5 to 3.0 Differences 14 15 The concept of tasks has been added to the OpenMP execution model (see Section 1.2.4 on page 8 and Section 1.3 on page 14). 16 17 • The task construct (see Section 2.11 on page 113) has been added, which provides 18 19 • The taskwait construct (see Section 2.12.4 on page 125) has been added, which 20 21 22 • The OpenMP memory model now covers atomicity of memory accesses (see 23 24 25 26 27 28 29 • In Version 2.5, there was a single copy of the nest-var, dyn-var, nthreads-var and 30 31 32 • The definition of active parallel region has been changed: in Version 3.0 a 33 34 • The rules for determining the number of threads used in a parallel region have 35 36 • In Version 3.0, the assignment of iterations to threads in a loop construct with a a mechanism for creating tasks explicitly. causes a task to wait for all its child tasks to complete. Section 1.4.1 on page 17). The description of the behavior of volatile in terms of flush was removed. run-sched-var internal control variables (ICVs) for the whole program. In Version 3.0, there is one copy of these ICVs per task (see Section 2.3 on page 34). 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 189, Section 3.2.7 on page 197 and Section 3.2.10 on page 200). parallel region is active if it is executed by a team consisting of more than one thread (see Section 1.2.2 on page 2). been modified (see Section 2.5.1 on page 47). static schedule kind is deterministic (see Section 2.7.1 on page 53). Appendix E Features History 305 1 2 3 • In Version 3.0, a loop construct may be associated with more than one perfectly 4 5 • Random access iterators, and variables of unsigned integer type, may now be used as 6 7 8 • The schedule kind auto has been added, which gives the implementation the nested loop. The number of associated loops may be controlled by the collapse clause (see Section 2.7.1 on page 53). loop iterators in loops associated with a loop construct (see Section 2.7.1 on page 53). freedom to choose any possible mapping of iterations in a loop construct to threads in the team (see Section 2.7.1 on page 53). 9 10 • Fortran assumed-size arrays now have predetermined data-sharing attributes (see 11 12 • In Fortran, firstprivate is now permitted as an argument to the default 13 14 15 16 17 • For list items in the private clause, implementations are no longer permitted to use 18 19 20 21 22 • In Version 3.0, Fortran allocatable arrays may appear in private, 23 24 • In Version 3.0, static class members variables may appear in a threadprivate 25 26 27 28 • Version 3.0 makes clear where, and with which arguments, constructors and 29 30 31 • The runtime library routines omp_set_schedule and omp_get_schedule 32 33 34 35 36 • The thread-limit-var ICV has been added, which controls the maximum number of 37 38 39 40 • The max-active-levels-var ICV has been added, which controls the number of nested Section 2.14.1.1 on page 146). clause (see Section 2.14.3.1 on page 156). 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 2.14.3.3 on page 159). firstprivate, lastprivate, reduction, copyin and copyprivate clauses. (see Section 2.14.2 on page 150, Section 2.14.3.3 on page 159, Section 2.14.3.4 on page 162, Section 2.14.3.5 on page 164, Section 2.14.3.6 on page 167, Section 2.14.4.1 on page 173 and Section 2.14.4.2 on page 175). directive (see Section 2.14.2 on page 150). destructors of private and threadprivate class type variables are called (see Section 2.14.2 on page 150, Section 2.14.3.3 on page 159, Section 2.14.3.4 on page 162, Section 2.14.4.1 on page 173 and Section 2.14.4.2 on page 175) have been added; these routines respectively set and retrieve the value of the run-sched-var ICV (see Section 3.2.12 on page 203 and Section 3.2.13 on page 205). 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 35, Section 3.2.14 on page 206 and Section 4.10 on page 246). 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 306 OpenMP API • Version 4.0 - July 2013 1 2 3 with the omp_get_max_active_levels runtime library routine (see Section 2.3.1 on page 35, Section 3.2.15 on page 207, Section 3.2.16 on page 209 and Section 4.9 on page 245). 4 5 6 7 • The stacksize-var ICV has been added, which controls the stack size for threads that the OpenMP implementation creates. The value of this ICV can be set with the OMP_STACKSIZE environment variable (see Section 2.3.1 on page 35 and Section 4.7 on page 244). 8 9 10 • The wait-policy-var ICV has been added, which controls the desired behavior of 11 12 13 • The omp_get_level runtime library routine has been added, which returns the 14 15 16 • The omp_get_ancestor_thread_num runtime library routine has been added, 17 18 19 • The omp_get_team_size runtime library routine has been added, which returns, 20 21 22 • The omp_get_active_level runtime library routine has been added, which 23 24 • In Version 3.0, locks are owned by tasks, not by threads (see Section 3.3 on page waiting threads. The value of this ICV can be set with the OMP_WAIT_POLICY environment variable (see Section 2.3.1 on page 35 and Section 4.8 on page 245). number of nested parallel regions enclosing the task that contains the call (see Section 3.2.17 on page 210). which returns, for a given nested level of the current thread, the thread number of the ancestor (see Section 3.2.18 on page 211). for a given nested level of the current thread, the size of the thread team to which the ancestor belongs (see Section 3.2.19 on page 212). returns the number of nested, active parallel regions enclosing the task that contains the call (see Section 3.2.20 on page 214). 224). Appendix E Features History 307 308 OpenMP API • Version 4.0 - July 2013 Index Symbols _OPENMP macro, 2-32 A array sections, 2-42 atomic, 2-127 atomic construct, 8-300 attributes, data-sharing, 2-146 auto, 2-57 B barrier, 2-123 C cancel, 2-140 cancellation constructs cancel, 2-140 cancellation point, 2-143 cancellation point, 2-143 capture, atomic, 2-127 clauses collapse, 2-55 copyin, 2-173 copyprivate, 2-175 data-sharing, 2-155 default, 2-156 depend, 2-116 firstprivate, 2-162 lastprivate, 2-164 map, 2-177 private, 2-159 reduction, 2-167 schedule, 2-56 shared, 2-157 collapse, 2-55 compliance, 1-21 conditional compilation, 2-32 constructs atomic, 2-127 barrier, 2-123 cancel, 2-140 cancellation point, 2-143 critical, 2-122 declare simd, 2-72 declare target, 2-83 distribute, 2-88 distribute parallel do, 2-92 distribute parallel do simd, 2-94 distribute parallel for, 2-92 distribute parallel for simd, 2-94 distribute parallel loop, 2-92 distribute simd, 2-91 do, Fortran, 2-54 flush, 2-134 for, C/C++, 2-54 loop, 2-53 Loop SIMD, 2-76 master, 2-120 ordered, 2-138 parallel, 2-44 parallel for, C/C++, 2-95 parallel sections, 2-97 parallel workshare, Fortran, 2-99 sections, 2-60 simd, 2-68 Index 309 single, 2-63 target, 2-79 target data, 2-77 target teams, 2-101 target teams distribute, 2-102, 2-105 target update, 2-81 task, 2-113 taskgroup, 2-126 taskwait, 2-125 taskyield, 2-117 teams, 2-86 teams distribute, 2-102 workshare, 2-65 worksharing, 2-53 copyin, 2-173 copyprivate, 2-175 critical, 2-122 dynamic, 2-57 dynamic thread adjustment, 8-299 E environment variables, 4-237 modifying ICV’s, 2-36 OMP_CANCELLATION, 4-246 OMP_DEFAULT_DEVICE, 4-248 OMP_DISPLAY_ENV, 4-247 OMP_DYNAMIC, 4-240 OMP_MAX_ACTIVE_LEVELS, 4-245 OMP_NESTED, 4-243 OMP_NUM_THREADS, 4-239 OMP_SCHEDULE, 4-238 OMP_STACKSIZE, 4-244 OMP_THREAD_LIMIT, 4-246 OMP_WAIT_POLICY, 4-245 execution model, 1-14 D data sharing, 2-146 data-sharing clauses, 2-155 declare reduction, 2-180 declare simd construct, 2-72 declare target, 2-83 default, 2-156 depend, 2-116 device constructs declare target, 2-83 distribute, 2-88 target, 2-79 target data, 2-77 target update, 2-81 teams, 2-86 device data environments, 1-18 directives, 2-25 format, 2-26 threadprivate, 2-150 see also constructs distribute, 2-88 distribute parallel do, 2-92 distribute parallel do simd, 2-94 distribute parallel for, 2-92 distribute parallel for simd, 2-94 distribute simd, 2-91 do simd, 2-76 do, Fortran, 2-54 Index-310 OpenMP API • Version 4.0 - July 2013 F firstprivate, 2-162 flush, 2-134 flush operation, 1-19 for simd, 2-76 for, C/C++, 2-54 G glossary, 1-2 grammar rules, 6-266 guided, 2-57 H header files, 3-188, 7-287 I ICVs (internal control variables), 2-34 implementation, 8-299 include files, 3-188, 7-287 internal control variables, 8-299 internal control variables (ICVs), 2-34 L lastprivate, 2-164 loop directive, 8-300 loop SIMD construct, 2-76 loop, scheduling, 2-59 M map, 2-177 master, 2-120 memory model, 1-17, 8-299 model execution, 1-14 memory, 1-17 N nested parallelism, 1-15, 2-34, 3-200 nesting, 2-186 number of threads, 2-47 O OMP_CANCELLATION, 4-246 OMP_DEFAULT_DEVICE, 4-248 omp_destroy_lock, 3-227 omp_destroy_nest_lock, 3-227 OMP_DISPLAY_ENV, 4-247 OMP_DYNAMIC, 4-240, 8-301 omp_get_active_level, 3-214 omp_get_ancestor_thread_num, 3-211 omp_get_cancellation, 3-199 omp_get_default_device, 3-219 omp_get_dynamic, 3-198 omp_get_level, 3-210 omp_get_max_active_levels, 3-209, 8-301 omp_get_max_threads, 3-192 omp_get_nested, 3-201 omp_get_num_devices, 3-220 omp_get_num_procs, 3-195 omp_get_num_teams, 3-221 omp_get_num_threads, 3-191 omp_get_proc_bind, 3-216 omp_get_schedule, 3-205 omp_get_team_num, 3-222 omp_get_team_size, 3-212 omp_get_thread_limit, 3-206 omp_get_thread_num, 3-193 omp_get_wtick, 3-234 omp_get_wtime, 3-233 omp_in_final, 3-215 omp_in_parallel, 3-196 omp_init_lock, 3-226 omp_init_nest_lock, 3-226 omp_is_initial_device, 3-223 omp_lock_kind, 3-225 omp_lock_t, 3-225 OMP_MAX_ACTIVE_LEVELS, 4-245, 8-301 omp_nest_lock_kind, 3-225 omp_nest_lock_t, 3-225 OMP_NESTED, 4-243, 8-301 OMP_NUM_THREADS, 4-239, 8-301 OMP_PLACES, 4-241, 8-301 OMP_PROC_BIND, 8-301 OMP_SCHEDULE, 4-238, 8-301 omp_set_default_device, 3-218 omp_set_dynamic, 3-197 omp_set_lock, 3-228 omp_set_max_active_levels, 3-207, 8-301 omp_set_nest_lock, 3-228 omp_set_nested, 3-200 omp_set_num_threads, 3-189, 8-300 omp_set_schedule, 3-203, 8-300 OMP_STACKSIZE, 4-244, 8-301 omp_test_lock, 3-231 omp_test_nest_lock, 3-231 OMP_THREAD_LIMIT, 4-246, 8-301 omp_unset_lock, 3-229 omp_unset_nest_lock, 3-229 OMP_WAIT_POLICY, 4-245, 8-301 OpenMP compliance, 1-21 features history, 9-303 implementation, 8-299 ordered, 2-138 P parallel, 2-44 parallel do, 2-96 parallel do simd, 2-100 parallel for simd, 2-100 parallel for, C/C++, 2-95 parallel loop SIMD construct, 2-100 parallel sections, 2-97 parallel workshare, Fortran, 2-99 Index-311 pragmas see constructs private, 2-159 R read, atomic, 2-127 reduction, 2-167 references, 1-22 regions, nesting, 2-186 runtime, 2-58 runtime library interfaces and prototypes, 3-188 runtime library definitions, 8-302 S schedule, 2-56 scheduling loop, 2-59 tasks, 2-118 sections, 2-60 sections construct, 8-300 shared, 2-157 shared clause, 8-302 simd, 2-68 simd construct, 2-68 SIMD lanes, 1-15 SIMD loop, 2-68 SIMD loop construct, 2-76 SIMD parallel loop construct, 2-100 single, 2-63 single construct, 8-300 static, 2-57 stubs for runtime library routines C/C++, 5-250 Fortran, 5-257 synchronization, locks constructs, 2-120 routines, 3-224 T target, target target target Index-312 2-79 data, 2-77 teams, 2-101 teams distribute, 2-102, 2-105 OpenMP API • Version 4.0 - July 2013 target update, 2-81 task scheduling, 2-118 task, 2-113 taskgroup, 2-126 tasking, 2-113 taskwait, 2-125 taskyield, 2-117 teams, 2-86 teams distribute, 2-102 terminology, 1-2 thread affinity policy, 8-302 threadprivate, 2-150, 8-302 timer, 3-233 timing routines, 3-233 U update, atomic, 2-127 V variables, environment, 4-237 W wall clock timer, 3-233 website www.openmp.org workshare, 2-65 worksharing constructs, 2-53 parallel, 2-95 scheduling, 2-59 write, atomic, 2-127