1\input texinfo @c -*-texinfo-*- 2 3@c %**start of header 4@setfilename libgomp.info 5@settitle GNU libgomp 6@c %**end of header 7 8 9@copying 10Copyright @copyright{} 2006-2016 Free Software Foundation, Inc. 11 12Permission is granted to copy, distribute and/or modify this document 13under the terms of the GNU Free Documentation License, Version 1.3 or 14any later version published by the Free Software Foundation; with the 15Invariant Sections being ``Funding Free Software'', the Front-Cover 16texts being (a) (see below), and with the Back-Cover Texts being (b) 17(see below). A copy of the license is included in the section entitled 18``GNU Free Documentation License''. 19 20(a) The FSF's Front-Cover Text is: 21 22 A GNU Manual 23 24(b) The FSF's Back-Cover Text is: 25 26 You have freedom to copy and modify this GNU Manual, like GNU 27 software. Copies published by the Free Software Foundation raise 28 funds for GNU development. 29@end copying 30 31@ifinfo 32@dircategory GNU Libraries 33@direntry 34* libgomp: (libgomp). GNU Offloading and Multi Processing Runtime Library. 35@end direntry 36 37This manual documents libgomp, the GNU Offloading and Multi Processing 38Runtime library. This is the GNU implementation of the OpenMP and 39OpenACC APIs for parallel and accelerator programming in C/C++ and 40Fortran. 41 42Published by the Free Software Foundation 4351 Franklin Street, Fifth Floor 44Boston, MA 02110-1301 USA 45 46@insertcopying 47@end ifinfo 48 49 50@setchapternewpage odd 51 52@titlepage 53@title GNU Offloading and Multi Processing Runtime Library 54@subtitle The GNU OpenMP and OpenACC Implementation 55@page 56@vskip 0pt plus 1filll 57@comment For the @value{version-GCC} Version* 58@sp 1 59Published by the Free Software Foundation @* 6051 Franklin Street, Fifth Floor@* 61Boston, MA 02110-1301, USA@* 62@sp 1 63@insertcopying 64@end titlepage 65 66@summarycontents 67@contents 68@page 69 70 71@node Top 72@top Introduction 73@cindex Introduction 74 75This manual documents the usage of libgomp, the GNU Offloading and 76Multi Processing Runtime Library. This includes the GNU 77implementation of the @uref{http://www.openmp.org, OpenMP} Application 78Programming Interface (API) for multi-platform shared-memory parallel 79programming in C/C++ and Fortran, and the GNU implementation of the 80@uref{http://www.openacc.org/, OpenACC} Application Programming 81Interface (API) for offloading of code to accelerator devices in C/C++ 82and Fortran. 83 84Originally, libgomp implemented the GNU OpenMP Runtime Library. Based 85on this, support for OpenACC and offloading (both OpenACC and OpenMP 864's target construct) has been added later on, and the library's name 87changed to GNU Offloading and Multi Processing Runtime Library. 88 89 90 91@comment 92@comment When you add a new menu item, please keep the right hand 93@comment aligned to the same column. Do not use tabs. This provides 94@comment better formatting. 95@comment 96@menu 97* Enabling OpenMP:: How to enable OpenMP for your applications. 98* Runtime Library Routines:: The OpenMP runtime application programming 99 interface. 100* Environment Variables:: Influencing runtime behavior with environment 101 variables. 102* Enabling OpenACC:: How to enable OpenACC for your 103 applications. 104* OpenACC Runtime Library Routines:: The OpenACC runtime application 105 programming interface. 106* OpenACC Environment Variables:: Influencing OpenACC runtime behavior with 107 environment variables. 108* CUDA Streams Usage:: Notes on the implementation of 109 asynchronous operations. 110* OpenACC Library Interoperability:: OpenACC library interoperability with the 111 NVIDIA CUBLAS library. 112* The libgomp ABI:: Notes on the external ABI presented by libgomp. 113* Reporting Bugs:: How to report bugs in the GNU Offloading and 114 Multi Processing Runtime Library. 115* Copying:: GNU general public license says 116 how you can copy and share libgomp. 117* GNU Free Documentation License:: 118 How you can copy and share this manual. 119* Funding:: How to help assure continued work for free 120 software. 121* Library Index:: Index of this documentation. 122@end menu 123 124 125@c --------------------------------------------------------------------- 126@c Enabling OpenMP 127@c --------------------------------------------------------------------- 128 129@node Enabling OpenMP 130@chapter Enabling OpenMP 131 132To activate the OpenMP extensions for C/C++ and Fortran, the compile-time 133flag @command{-fopenmp} must be specified. This enables the OpenMP directive 134@code{#pragma omp} in C/C++ and @code{!$omp} directives in free form, 135@code{c$omp}, @code{*$omp} and @code{!$omp} directives in fixed form, 136@code{!$} conditional compilation sentinels in free form and @code{c$}, 137@code{*$} and @code{!$} sentinels in fixed form, for Fortran. The flag also 138arranges for automatic linking of the OpenMP runtime library 139(@ref{Runtime Library Routines}). 140 141A complete description of all OpenMP directives accepted may be found in 142the @uref{http://www.openmp.org, OpenMP Application Program Interface} manual, 143version 4.5. 144 145 146@c --------------------------------------------------------------------- 147@c Runtime Library Routines 148@c --------------------------------------------------------------------- 149 150@node Runtime Library Routines 151@chapter Runtime Library Routines 152 153The runtime routines described here are defined by Section 3 of the OpenMP 154specification in version 4.5. The routines are structured in following 155three parts: 156 157@menu 158Control threads, processors and the parallel environment. They have C 159linkage, and do not throw exceptions. 160 161* omp_get_active_level:: Number of active parallel regions 162* omp_get_ancestor_thread_num:: Ancestor thread ID 163* omp_get_cancellation:: Whether cancellation support is enabled 164* omp_get_default_device:: Get the default device for target regions 165* omp_get_dynamic:: Dynamic teams setting 166* omp_get_level:: Number of parallel regions 167* omp_get_max_active_levels:: Maximum number of active regions 168* omp_get_max_task_priority:: Maximum task priority value that can be set 169* omp_get_max_threads:: Maximum number of threads of parallel region 170* omp_get_nested:: Nested parallel regions 171* omp_get_num_devices:: Number of target devices 172* omp_get_num_procs:: Number of processors online 173* omp_get_num_teams:: Number of teams 174* omp_get_num_threads:: Size of the active team 175* omp_get_proc_bind:: Whether theads may be moved between CPUs 176* omp_get_schedule:: Obtain the runtime scheduling method 177* omp_get_team_num:: Get team number 178* omp_get_team_size:: Number of threads in a team 179* omp_get_thread_limit:: Maximum number of threads 180* omp_get_thread_num:: Current thread ID 181* omp_in_parallel:: Whether a parallel region is active 182* omp_in_final:: Whether in final or included task region 183* omp_is_initial_device:: Whether executing on the host device 184* omp_set_default_device:: Set the default device for target regions 185* omp_set_dynamic:: Enable/disable dynamic teams 186* omp_set_max_active_levels:: Limits the number of active parallel regions 187* omp_set_nested:: Enable/disable nested parallel regions 188* omp_set_num_threads:: Set upper team size limit 189* omp_set_schedule:: Set the runtime scheduling method 190 191Initialize, set, test, unset and destroy simple and nested locks. 192 193* omp_init_lock:: Initialize simple lock 194* omp_set_lock:: Wait for and set simple lock 195* omp_test_lock:: Test and set simple lock if available 196* omp_unset_lock:: Unset simple lock 197* omp_destroy_lock:: Destroy simple lock 198* omp_init_nest_lock:: Initialize nested lock 199* omp_set_nest_lock:: Wait for and set simple lock 200* omp_test_nest_lock:: Test and set nested lock if available 201* omp_unset_nest_lock:: Unset nested lock 202* omp_destroy_nest_lock:: Destroy nested lock 203 204Portable, thread-based, wall clock timer. 205 206* omp_get_wtick:: Get timer precision. 207* omp_get_wtime:: Elapsed wall clock time. 208@end menu 209 210 211 212@node omp_get_active_level 213@section @code{omp_get_active_level} -- Number of parallel regions 214@table @asis 215@item @emph{Description}: 216This function returns the nesting level for the active parallel blocks, 217which enclose the calling call. 218 219@item @emph{C/C++} 220@multitable @columnfractions .20 .80 221@item @emph{Prototype}: @tab @code{int omp_get_active_level(void);} 222@end multitable 223 224@item @emph{Fortran}: 225@multitable @columnfractions .20 .80 226@item @emph{Interface}: @tab @code{integer function omp_get_active_level()} 227@end multitable 228 229@item @emph{See also}: 230@ref{omp_get_level}, @ref{omp_get_max_active_levels}, @ref{omp_set_max_active_levels} 231 232@item @emph{Reference}: 233@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.20. 234@end table 235 236 237 238@node omp_get_ancestor_thread_num 239@section @code{omp_get_ancestor_thread_num} -- Ancestor thread ID 240@table @asis 241@item @emph{Description}: 242This function returns the thread identification number for the given 243nesting level of the current thread. For values of @var{level} outside 244zero to @code{omp_get_level} -1 is returned; if @var{level} is 245@code{omp_get_level} the result is identical to @code{omp_get_thread_num}. 246 247@item @emph{C/C++} 248@multitable @columnfractions .20 .80 249@item @emph{Prototype}: @tab @code{int omp_get_ancestor_thread_num(int level);} 250@end multitable 251 252@item @emph{Fortran}: 253@multitable @columnfractions .20 .80 254@item @emph{Interface}: @tab @code{integer function omp_get_ancestor_thread_num(level)} 255@item @tab @code{integer level} 256@end multitable 257 258@item @emph{See also}: 259@ref{omp_get_level}, @ref{omp_get_thread_num}, @ref{omp_get_team_size} 260 261@item @emph{Reference}: 262@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.18. 263@end table 264 265 266 267@node omp_get_cancellation 268@section @code{omp_get_cancellation} -- Whether cancellation support is enabled 269@table @asis 270@item @emph{Description}: 271This function returns @code{true} if cancellation is activated, @code{false} 272otherwise. Here, @code{true} and @code{false} represent their language-specific 273counterparts. Unless @env{OMP_CANCELLATION} is set true, cancellations are 274deactivated. 275 276@item @emph{C/C++}: 277@multitable @columnfractions .20 .80 278@item @emph{Prototype}: @tab @code{int omp_get_cancellation(void);} 279@end multitable 280 281@item @emph{Fortran}: 282@multitable @columnfractions .20 .80 283@item @emph{Interface}: @tab @code{logical function omp_get_cancellation()} 284@end multitable 285 286@item @emph{See also}: 287@ref{OMP_CANCELLATION} 288 289@item @emph{Reference}: 290@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.9. 291@end table 292 293 294 295@node omp_get_default_device 296@section @code{omp_get_default_device} -- Get the default device for target regions 297@table @asis 298@item @emph{Description}: 299Get the default device for target regions without device clause. 300 301@item @emph{C/C++}: 302@multitable @columnfractions .20 .80 303@item @emph{Prototype}: @tab @code{int omp_get_default_device(void);} 304@end multitable 305 306@item @emph{Fortran}: 307@multitable @columnfractions .20 .80 308@item @emph{Interface}: @tab @code{integer function omp_get_default_device()} 309@end multitable 310 311@item @emph{See also}: 312@ref{OMP_DEFAULT_DEVICE}, @ref{omp_set_default_device} 313 314@item @emph{Reference}: 315@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.30. 316@end table 317 318 319 320@node omp_get_dynamic 321@section @code{omp_get_dynamic} -- Dynamic teams setting 322@table @asis 323@item @emph{Description}: 324This function returns @code{true} if enabled, @code{false} otherwise. 325Here, @code{true} and @code{false} represent their language-specific 326counterparts. 327 328The dynamic team setting may be initialized at startup by the 329@env{OMP_DYNAMIC} environment variable or at runtime using 330@code{omp_set_dynamic}. If undefined, dynamic adjustment is 331disabled by default. 332 333@item @emph{C/C++}: 334@multitable @columnfractions .20 .80 335@item @emph{Prototype}: @tab @code{int omp_get_dynamic(void);} 336@end multitable 337 338@item @emph{Fortran}: 339@multitable @columnfractions .20 .80 340@item @emph{Interface}: @tab @code{logical function omp_get_dynamic()} 341@end multitable 342 343@item @emph{See also}: 344@ref{omp_set_dynamic}, @ref{OMP_DYNAMIC} 345 346@item @emph{Reference}: 347@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.8. 348@end table 349 350 351 352@node omp_get_level 353@section @code{omp_get_level} -- Obtain the current nesting level 354@table @asis 355@item @emph{Description}: 356This function returns the nesting level for the parallel blocks, 357which enclose the calling call. 358 359@item @emph{C/C++} 360@multitable @columnfractions .20 .80 361@item @emph{Prototype}: @tab @code{int omp_get_level(void);} 362@end multitable 363 364@item @emph{Fortran}: 365@multitable @columnfractions .20 .80 366@item @emph{Interface}: @tab @code{integer function omp_level()} 367@end multitable 368 369@item @emph{See also}: 370@ref{omp_get_active_level} 371 372@item @emph{Reference}: 373@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.17. 374@end table 375 376 377 378@node omp_get_max_active_levels 379@section @code{omp_get_max_active_levels} -- Maximum number of active regions 380@table @asis 381@item @emph{Description}: 382This function obtains the maximum allowed number of nested, active parallel regions. 383 384@item @emph{C/C++} 385@multitable @columnfractions .20 .80 386@item @emph{Prototype}: @tab @code{int omp_get_max_active_levels(void);} 387@end multitable 388 389@item @emph{Fortran}: 390@multitable @columnfractions .20 .80 391@item @emph{Interface}: @tab @code{integer function omp_get_max_active_levels()} 392@end multitable 393 394@item @emph{See also}: 395@ref{omp_set_max_active_levels}, @ref{omp_get_active_level} 396 397@item @emph{Reference}: 398@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.16. 399@end table 400 401 402@node omp_get_max_task_priority 403@section @code{omp_get_max_task_priority} -- Maximum priority value 404that can be set for tasks. 405@table @asis 406@item @emph{Description}: 407This function obtains the maximum allowed priority number for tasks. 408 409@item @emph{C/C++} 410@multitable @columnfractions .20 .80 411@item @emph{Prototype}: @tab @code{int omp_get_max_task_priority(void);} 412@end multitable 413 414@item @emph{Fortran}: 415@multitable @columnfractions .20 .80 416@item @emph{Interface}: @tab @code{integer function omp_get_max_task_priority()} 417@end multitable 418 419@item @emph{Reference}: 420@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.29. 421@end table 422 423 424@node omp_get_max_threads 425@section @code{omp_get_max_threads} -- Maximum number of threads of parallel region 426@table @asis 427@item @emph{Description}: 428Return the maximum number of threads used for the current parallel region 429that does not use the clause @code{num_threads}. 430 431@item @emph{C/C++}: 432@multitable @columnfractions .20 .80 433@item @emph{Prototype}: @tab @code{int omp_get_max_threads(void);} 434@end multitable 435 436@item @emph{Fortran}: 437@multitable @columnfractions .20 .80 438@item @emph{Interface}: @tab @code{integer function omp_get_max_threads()} 439@end multitable 440 441@item @emph{See also}: 442@ref{omp_set_num_threads}, @ref{omp_set_dynamic}, @ref{omp_get_thread_limit} 443 444@item @emph{Reference}: 445@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.3. 446@end table 447 448 449 450@node omp_get_nested 451@section @code{omp_get_nested} -- Nested parallel regions 452@table @asis 453@item @emph{Description}: 454This function returns @code{true} if nested parallel regions are 455enabled, @code{false} otherwise. Here, @code{true} and @code{false} 456represent their language-specific counterparts. 457 458Nested parallel regions may be initialized at startup by the 459@env{OMP_NESTED} environment variable or at runtime using 460@code{omp_set_nested}. If undefined, nested parallel regions are 461disabled by default. 462 463@item @emph{C/C++}: 464@multitable @columnfractions .20 .80 465@item @emph{Prototype}: @tab @code{int omp_get_nested(void);} 466@end multitable 467 468@item @emph{Fortran}: 469@multitable @columnfractions .20 .80 470@item @emph{Interface}: @tab @code{logical function omp_get_nested()} 471@end multitable 472 473@item @emph{See also}: 474@ref{omp_set_nested}, @ref{OMP_NESTED} 475 476@item @emph{Reference}: 477@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.11. 478@end table 479 480 481 482@node omp_get_num_devices 483@section @code{omp_get_num_devices} -- Number of target devices 484@table @asis 485@item @emph{Description}: 486Returns the number of target devices. 487 488@item @emph{C/C++}: 489@multitable @columnfractions .20 .80 490@item @emph{Prototype}: @tab @code{int omp_get_num_devices(void);} 491@end multitable 492 493@item @emph{Fortran}: 494@multitable @columnfractions .20 .80 495@item @emph{Interface}: @tab @code{integer function omp_get_num_devices()} 496@end multitable 497 498@item @emph{Reference}: 499@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.31. 500@end table 501 502 503 504@node omp_get_num_procs 505@section @code{omp_get_num_procs} -- Number of processors online 506@table @asis 507@item @emph{Description}: 508Returns the number of processors online on that device. 509 510@item @emph{C/C++}: 511@multitable @columnfractions .20 .80 512@item @emph{Prototype}: @tab @code{int omp_get_num_procs(void);} 513@end multitable 514 515@item @emph{Fortran}: 516@multitable @columnfractions .20 .80 517@item @emph{Interface}: @tab @code{integer function omp_get_num_procs()} 518@end multitable 519 520@item @emph{Reference}: 521@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.5. 522@end table 523 524 525 526@node omp_get_num_teams 527@section @code{omp_get_num_teams} -- Number of teams 528@table @asis 529@item @emph{Description}: 530Returns the number of teams in the current team region. 531 532@item @emph{C/C++}: 533@multitable @columnfractions .20 .80 534@item @emph{Prototype}: @tab @code{int omp_get_num_teams(void);} 535@end multitable 536 537@item @emph{Fortran}: 538@multitable @columnfractions .20 .80 539@item @emph{Interface}: @tab @code{integer function omp_get_num_teams()} 540@end multitable 541 542@item @emph{Reference}: 543@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.32. 544@end table 545 546 547 548@node omp_get_num_threads 549@section @code{omp_get_num_threads} -- Size of the active team 550@table @asis 551@item @emph{Description}: 552Returns the number of threads in the current team. In a sequential section of 553the program @code{omp_get_num_threads} returns 1. 554 555The default team size may be initialized at startup by the 556@env{OMP_NUM_THREADS} environment variable. At runtime, the size 557of the current team may be set either by the @code{NUM_THREADS} 558clause or by @code{omp_set_num_threads}. If none of the above were 559used to define a specific value and @env{OMP_DYNAMIC} is disabled, 560one thread per CPU online is used. 561 562@item @emph{C/C++}: 563@multitable @columnfractions .20 .80 564@item @emph{Prototype}: @tab @code{int omp_get_num_threads(void);} 565@end multitable 566 567@item @emph{Fortran}: 568@multitable @columnfractions .20 .80 569@item @emph{Interface}: @tab @code{integer function omp_get_num_threads()} 570@end multitable 571 572@item @emph{See also}: 573@ref{omp_get_max_threads}, @ref{omp_set_num_threads}, @ref{OMP_NUM_THREADS} 574 575@item @emph{Reference}: 576@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.2. 577@end table 578 579 580 581@node omp_get_proc_bind 582@section @code{omp_get_proc_bind} -- Whether theads may be moved between CPUs 583@table @asis 584@item @emph{Description}: 585This functions returns the currently active thread affinity policy, which is 586set via @env{OMP_PROC_BIND}. Possible values are @code{omp_proc_bind_false}, 587@code{omp_proc_bind_true}, @code{omp_proc_bind_master}, 588@code{omp_proc_bind_close} and @code{omp_proc_bind_spread}. 589 590@item @emph{C/C++}: 591@multitable @columnfractions .20 .80 592@item @emph{Prototype}: @tab @code{omp_proc_bind_t omp_get_proc_bind(void);} 593@end multitable 594 595@item @emph{Fortran}: 596@multitable @columnfractions .20 .80 597@item @emph{Interface}: @tab @code{integer(kind=omp_proc_bind_kind) function omp_get_proc_bind()} 598@end multitable 599 600@item @emph{See also}: 601@ref{OMP_PROC_BIND}, @ref{OMP_PLACES}, @ref{GOMP_CPU_AFFINITY}, 602 603@item @emph{Reference}: 604@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.22. 605@end table 606 607 608 609@node omp_get_schedule 610@section @code{omp_get_schedule} -- Obtain the runtime scheduling method 611@table @asis 612@item @emph{Description}: 613Obtain the runtime scheduling method. The @var{kind} argument will be 614set to the value @code{omp_sched_static}, @code{omp_sched_dynamic}, 615@code{omp_sched_guided} or @code{omp_sched_auto}. The second argument, 616@var{chunk_size}, is set to the chunk size. 617 618@item @emph{C/C++} 619@multitable @columnfractions .20 .80 620@item @emph{Prototype}: @tab @code{void omp_get_schedule(omp_sched_t *kind, int *chunk_size);} 621@end multitable 622 623@item @emph{Fortran}: 624@multitable @columnfractions .20 .80 625@item @emph{Interface}: @tab @code{subroutine omp_get_schedule(kind, chunk_size)} 626@item @tab @code{integer(kind=omp_sched_kind) kind} 627@item @tab @code{integer chunk_size} 628@end multitable 629 630@item @emph{See also}: 631@ref{omp_set_schedule}, @ref{OMP_SCHEDULE} 632 633@item @emph{Reference}: 634@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.13. 635@end table 636 637 638 639@node omp_get_team_num 640@section @code{omp_get_team_num} -- Get team number 641@table @asis 642@item @emph{Description}: 643Returns the team number of the calling thread. 644 645@item @emph{C/C++}: 646@multitable @columnfractions .20 .80 647@item @emph{Prototype}: @tab @code{int omp_get_team_num(void);} 648@end multitable 649 650@item @emph{Fortran}: 651@multitable @columnfractions .20 .80 652@item @emph{Interface}: @tab @code{integer function omp_get_team_num()} 653@end multitable 654 655@item @emph{Reference}: 656@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.33. 657@end table 658 659 660 661@node omp_get_team_size 662@section @code{omp_get_team_size} -- Number of threads in a team 663@table @asis 664@item @emph{Description}: 665This function returns the number of threads in a thread team to which 666either the current thread or its ancestor belongs. For values of @var{level} 667outside zero to @code{omp_get_level}, -1 is returned; if @var{level} is zero, 6681 is returned, and for @code{omp_get_level}, the result is identical 669to @code{omp_get_num_threads}. 670 671@item @emph{C/C++}: 672@multitable @columnfractions .20 .80 673@item @emph{Prototype}: @tab @code{int omp_get_team_size(int level);} 674@end multitable 675 676@item @emph{Fortran}: 677@multitable @columnfractions .20 .80 678@item @emph{Interface}: @tab @code{integer function omp_get_team_size(level)} 679@item @tab @code{integer level} 680@end multitable 681 682@item @emph{See also}: 683@ref{omp_get_num_threads}, @ref{omp_get_level}, @ref{omp_get_ancestor_thread_num} 684 685@item @emph{Reference}: 686@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.19. 687@end table 688 689 690 691@node omp_get_thread_limit 692@section @code{omp_get_thread_limit} -- Maximum number of threads 693@table @asis 694@item @emph{Description}: 695Return the maximum number of threads of the program. 696 697@item @emph{C/C++}: 698@multitable @columnfractions .20 .80 699@item @emph{Prototype}: @tab @code{int omp_get_thread_limit(void);} 700@end multitable 701 702@item @emph{Fortran}: 703@multitable @columnfractions .20 .80 704@item @emph{Interface}: @tab @code{integer function omp_get_thread_limit()} 705@end multitable 706 707@item @emph{See also}: 708@ref{omp_get_max_threads}, @ref{OMP_THREAD_LIMIT} 709 710@item @emph{Reference}: 711@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.14. 712@end table 713 714 715 716@node omp_get_thread_num 717@section @code{omp_get_thread_num} -- Current thread ID 718@table @asis 719@item @emph{Description}: 720Returns a unique thread identification number within the current team. 721In a sequential parts of the program, @code{omp_get_thread_num} 722always returns 0. In parallel regions the return value varies 723from 0 to @code{omp_get_num_threads}-1 inclusive. The return 724value of the master thread of a team is always 0. 725 726@item @emph{C/C++}: 727@multitable @columnfractions .20 .80 728@item @emph{Prototype}: @tab @code{int omp_get_thread_num(void);} 729@end multitable 730 731@item @emph{Fortran}: 732@multitable @columnfractions .20 .80 733@item @emph{Interface}: @tab @code{integer function omp_get_thread_num()} 734@end multitable 735 736@item @emph{See also}: 737@ref{omp_get_num_threads}, @ref{omp_get_ancestor_thread_num} 738 739@item @emph{Reference}: 740@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.4. 741@end table 742 743 744 745@node omp_in_parallel 746@section @code{omp_in_parallel} -- Whether a parallel region is active 747@table @asis 748@item @emph{Description}: 749This function returns @code{true} if currently running in parallel, 750@code{false} otherwise. Here, @code{true} and @code{false} represent 751their language-specific counterparts. 752 753@item @emph{C/C++}: 754@multitable @columnfractions .20 .80 755@item @emph{Prototype}: @tab @code{int omp_in_parallel(void);} 756@end multitable 757 758@item @emph{Fortran}: 759@multitable @columnfractions .20 .80 760@item @emph{Interface}: @tab @code{logical function omp_in_parallel()} 761@end multitable 762 763@item @emph{Reference}: 764@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.6. 765@end table 766 767 768@node omp_in_final 769@section @code{omp_in_final} -- Whether in final or included task region 770@table @asis 771@item @emph{Description}: 772This function returns @code{true} if currently running in a final 773or included task region, @code{false} otherwise. Here, @code{true} 774and @code{false} represent their language-specific counterparts. 775 776@item @emph{C/C++}: 777@multitable @columnfractions .20 .80 778@item @emph{Prototype}: @tab @code{int omp_in_final(void);} 779@end multitable 780 781@item @emph{Fortran}: 782@multitable @columnfractions .20 .80 783@item @emph{Interface}: @tab @code{logical function omp_in_final()} 784@end multitable 785 786@item @emph{Reference}: 787@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.21. 788@end table 789 790 791 792@node omp_is_initial_device 793@section @code{omp_is_initial_device} -- Whether executing on the host device 794@table @asis 795@item @emph{Description}: 796This function returns @code{true} if currently running on the host device, 797@code{false} otherwise. Here, @code{true} and @code{false} represent 798their language-specific counterparts. 799 800@item @emph{C/C++}: 801@multitable @columnfractions .20 .80 802@item @emph{Prototype}: @tab @code{int omp_is_initial_device(void);} 803@end multitable 804 805@item @emph{Fortran}: 806@multitable @columnfractions .20 .80 807@item @emph{Interface}: @tab @code{logical function omp_is_initial_device()} 808@end multitable 809 810@item @emph{Reference}: 811@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.34. 812@end table 813 814 815 816@node omp_set_default_device 817@section @code{omp_set_default_device} -- Set the default device for target regions 818@table @asis 819@item @emph{Description}: 820Set the default device for target regions without device clause. The argument 821shall be a nonnegative device number. 822 823@item @emph{C/C++}: 824@multitable @columnfractions .20 .80 825@item @emph{Prototype}: @tab @code{void omp_set_default_device(int device_num);} 826@end multitable 827 828@item @emph{Fortran}: 829@multitable @columnfractions .20 .80 830@item @emph{Interface}: @tab @code{subroutine omp_set_default_device(device_num)} 831@item @tab @code{integer device_num} 832@end multitable 833 834@item @emph{See also}: 835@ref{OMP_DEFAULT_DEVICE}, @ref{omp_get_default_device} 836 837@item @emph{Reference}: 838@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.29. 839@end table 840 841 842 843@node omp_set_dynamic 844@section @code{omp_set_dynamic} -- Enable/disable dynamic teams 845@table @asis 846@item @emph{Description}: 847Enable or disable the dynamic adjustment of the number of threads 848within a team. The function takes the language-specific equivalent 849of @code{true} and @code{false}, where @code{true} enables dynamic 850adjustment of team sizes and @code{false} disables it. 851 852@item @emph{C/C++}: 853@multitable @columnfractions .20 .80 854@item @emph{Prototype}: @tab @code{void omp_set_dynamic(int dynamic_threads);} 855@end multitable 856 857@item @emph{Fortran}: 858@multitable @columnfractions .20 .80 859@item @emph{Interface}: @tab @code{subroutine omp_set_dynamic(dynamic_threads)} 860@item @tab @code{logical, intent(in) :: dynamic_threads} 861@end multitable 862 863@item @emph{See also}: 864@ref{OMP_DYNAMIC}, @ref{omp_get_dynamic} 865 866@item @emph{Reference}: 867@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.7. 868@end table 869 870 871 872@node omp_set_max_active_levels 873@section @code{omp_set_max_active_levels} -- Limits the number of active parallel regions 874@table @asis 875@item @emph{Description}: 876This function limits the maximum allowed number of nested, active 877parallel regions. 878 879@item @emph{C/C++} 880@multitable @columnfractions .20 .80 881@item @emph{Prototype}: @tab @code{void omp_set_max_active_levels(int max_levels);} 882@end multitable 883 884@item @emph{Fortran}: 885@multitable @columnfractions .20 .80 886@item @emph{Interface}: @tab @code{subroutine omp_set_max_active_levels(max_levels)} 887@item @tab @code{integer max_levels} 888@end multitable 889 890@item @emph{See also}: 891@ref{omp_get_max_active_levels}, @ref{omp_get_active_level} 892 893@item @emph{Reference}: 894@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.15. 895@end table 896 897 898 899@node omp_set_nested 900@section @code{omp_set_nested} -- Enable/disable nested parallel regions 901@table @asis 902@item @emph{Description}: 903Enable or disable nested parallel regions, i.e., whether team members 904are allowed to create new teams. The function takes the language-specific 905equivalent of @code{true} and @code{false}, where @code{true} enables 906dynamic adjustment of team sizes and @code{false} disables it. 907 908@item @emph{C/C++}: 909@multitable @columnfractions .20 .80 910@item @emph{Prototype}: @tab @code{void omp_set_nested(int nested);} 911@end multitable 912 913@item @emph{Fortran}: 914@multitable @columnfractions .20 .80 915@item @emph{Interface}: @tab @code{subroutine omp_set_nested(nested)} 916@item @tab @code{logical, intent(in) :: nested} 917@end multitable 918 919@item @emph{See also}: 920@ref{OMP_NESTED}, @ref{omp_get_nested} 921 922@item @emph{Reference}: 923@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.10. 924@end table 925 926 927 928@node omp_set_num_threads 929@section @code{omp_set_num_threads} -- Set upper team size limit 930@table @asis 931@item @emph{Description}: 932Specifies the number of threads used by default in subsequent parallel 933sections, if those do not specify a @code{num_threads} clause. The 934argument of @code{omp_set_num_threads} shall be a positive integer. 935 936@item @emph{C/C++}: 937@multitable @columnfractions .20 .80 938@item @emph{Prototype}: @tab @code{void omp_set_num_threads(int num_threads);} 939@end multitable 940 941@item @emph{Fortran}: 942@multitable @columnfractions .20 .80 943@item @emph{Interface}: @tab @code{subroutine omp_set_num_threads(num_threads)} 944@item @tab @code{integer, intent(in) :: num_threads} 945@end multitable 946 947@item @emph{See also}: 948@ref{OMP_NUM_THREADS}, @ref{omp_get_num_threads}, @ref{omp_get_max_threads} 949 950@item @emph{Reference}: 951@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.1. 952@end table 953 954 955 956@node omp_set_schedule 957@section @code{omp_set_schedule} -- Set the runtime scheduling method 958@table @asis 959@item @emph{Description}: 960Sets the runtime scheduling method. The @var{kind} argument can have the 961value @code{omp_sched_static}, @code{omp_sched_dynamic}, 962@code{omp_sched_guided} or @code{omp_sched_auto}. Except for 963@code{omp_sched_auto}, the chunk size is set to the value of 964@var{chunk_size} if positive, or to the default value if zero or negative. 965For @code{omp_sched_auto} the @var{chunk_size} argument is ignored. 966 967@item @emph{C/C++} 968@multitable @columnfractions .20 .80 969@item @emph{Prototype}: @tab @code{void omp_set_schedule(omp_sched_t kind, int chunk_size);} 970@end multitable 971 972@item @emph{Fortran}: 973@multitable @columnfractions .20 .80 974@item @emph{Interface}: @tab @code{subroutine omp_set_schedule(kind, chunk_size)} 975@item @tab @code{integer(kind=omp_sched_kind) kind} 976@item @tab @code{integer chunk_size} 977@end multitable 978 979@item @emph{See also}: 980@ref{omp_get_schedule} 981@ref{OMP_SCHEDULE} 982 983@item @emph{Reference}: 984@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.2.12. 985@end table 986 987 988 989@node omp_init_lock 990@section @code{omp_init_lock} -- Initialize simple lock 991@table @asis 992@item @emph{Description}: 993Initialize a simple lock. After initialization, the lock is in 994an unlocked state. 995 996@item @emph{C/C++}: 997@multitable @columnfractions .20 .80 998@item @emph{Prototype}: @tab @code{void omp_init_lock(omp_lock_t *lock);} 999@end multitable 1000 1001@item @emph{Fortran}: 1002@multitable @columnfractions .20 .80 1003@item @emph{Interface}: @tab @code{subroutine omp_init_lock(svar)} 1004@item @tab @code{integer(omp_lock_kind), intent(out) :: svar} 1005@end multitable 1006 1007@item @emph{See also}: 1008@ref{omp_destroy_lock} 1009 1010@item @emph{Reference}: 1011@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.3.1. 1012@end table 1013 1014 1015 1016@node omp_set_lock 1017@section @code{omp_set_lock} -- Wait for and set simple lock 1018@table @asis 1019@item @emph{Description}: 1020Before setting a simple lock, the lock variable must be initialized by 1021@code{omp_init_lock}. The calling thread is blocked until the lock 1022is available. If the lock is already held by the current thread, 1023a deadlock occurs. 1024 1025@item @emph{C/C++}: 1026@multitable @columnfractions .20 .80 1027@item @emph{Prototype}: @tab @code{void omp_set_lock(omp_lock_t *lock);} 1028@end multitable 1029 1030@item @emph{Fortran}: 1031@multitable @columnfractions .20 .80 1032@item @emph{Interface}: @tab @code{subroutine omp_set_lock(svar)} 1033@item @tab @code{integer(omp_lock_kind), intent(inout) :: svar} 1034@end multitable 1035 1036@item @emph{See also}: 1037@ref{omp_init_lock}, @ref{omp_test_lock}, @ref{omp_unset_lock} 1038 1039@item @emph{Reference}: 1040@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.3.4. 1041@end table 1042 1043 1044 1045@node omp_test_lock 1046@section @code{omp_test_lock} -- Test and set simple lock if available 1047@table @asis 1048@item @emph{Description}: 1049Before setting a simple lock, the lock variable must be initialized by 1050@code{omp_init_lock}. Contrary to @code{omp_set_lock}, @code{omp_test_lock} 1051does not block if the lock is not available. This function returns 1052@code{true} upon success, @code{false} otherwise. Here, @code{true} and 1053@code{false} represent their language-specific counterparts. 1054 1055@item @emph{C/C++}: 1056@multitable @columnfractions .20 .80 1057@item @emph{Prototype}: @tab @code{int omp_test_lock(omp_lock_t *lock);} 1058@end multitable 1059 1060@item @emph{Fortran}: 1061@multitable @columnfractions .20 .80 1062@item @emph{Interface}: @tab @code{logical function omp_test_lock(svar)} 1063@item @tab @code{integer(omp_lock_kind), intent(inout) :: svar} 1064@end multitable 1065 1066@item @emph{See also}: 1067@ref{omp_init_lock}, @ref{omp_set_lock}, @ref{omp_set_lock} 1068 1069@item @emph{Reference}: 1070@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.3.6. 1071@end table 1072 1073 1074 1075@node omp_unset_lock 1076@section @code{omp_unset_lock} -- Unset simple lock 1077@table @asis 1078@item @emph{Description}: 1079A simple lock about to be unset must have been locked by @code{omp_set_lock} 1080or @code{omp_test_lock} before. In addition, the lock must be held by the 1081thread calling @code{omp_unset_lock}. Then, the lock becomes unlocked. If one 1082or more threads attempted to set the lock before, one of them is chosen to, 1083again, set the lock to itself. 1084 1085@item @emph{C/C++}: 1086@multitable @columnfractions .20 .80 1087@item @emph{Prototype}: @tab @code{void omp_unset_lock(omp_lock_t *lock);} 1088@end multitable 1089 1090@item @emph{Fortran}: 1091@multitable @columnfractions .20 .80 1092@item @emph{Interface}: @tab @code{subroutine omp_unset_lock(svar)} 1093@item @tab @code{integer(omp_lock_kind), intent(inout) :: svar} 1094@end multitable 1095 1096@item @emph{See also}: 1097@ref{omp_set_lock}, @ref{omp_test_lock} 1098 1099@item @emph{Reference}: 1100@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.3.5. 1101@end table 1102 1103 1104 1105@node omp_destroy_lock 1106@section @code{omp_destroy_lock} -- Destroy simple lock 1107@table @asis 1108@item @emph{Description}: 1109Destroy a simple lock. In order to be destroyed, a simple lock must be 1110in the unlocked state. 1111 1112@item @emph{C/C++}: 1113@multitable @columnfractions .20 .80 1114@item @emph{Prototype}: @tab @code{void omp_destroy_lock(omp_lock_t *lock);} 1115@end multitable 1116 1117@item @emph{Fortran}: 1118@multitable @columnfractions .20 .80 1119@item @emph{Interface}: @tab @code{subroutine omp_destroy_lock(svar)} 1120@item @tab @code{integer(omp_lock_kind), intent(inout) :: svar} 1121@end multitable 1122 1123@item @emph{See also}: 1124@ref{omp_init_lock} 1125 1126@item @emph{Reference}: 1127@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.3.3. 1128@end table 1129 1130 1131 1132@node omp_init_nest_lock 1133@section @code{omp_init_nest_lock} -- Initialize nested lock 1134@table @asis 1135@item @emph{Description}: 1136Initialize a nested lock. After initialization, the lock is in 1137an unlocked state and the nesting count is set to zero. 1138 1139@item @emph{C/C++}: 1140@multitable @columnfractions .20 .80 1141@item @emph{Prototype}: @tab @code{void omp_init_nest_lock(omp_nest_lock_t *lock);} 1142@end multitable 1143 1144@item @emph{Fortran}: 1145@multitable @columnfractions .20 .80 1146@item @emph{Interface}: @tab @code{subroutine omp_init_nest_lock(nvar)} 1147@item @tab @code{integer(omp_nest_lock_kind), intent(out) :: nvar} 1148@end multitable 1149 1150@item @emph{See also}: 1151@ref{omp_destroy_nest_lock} 1152 1153@item @emph{Reference}: 1154@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.3.1. 1155@end table 1156 1157 1158@node omp_set_nest_lock 1159@section @code{omp_set_nest_lock} -- Wait for and set nested lock 1160@table @asis 1161@item @emph{Description}: 1162Before setting a nested lock, the lock variable must be initialized by 1163@code{omp_init_nest_lock}. The calling thread is blocked until the lock 1164is available. If the lock is already held by the current thread, the 1165nesting count for the lock is incremented. 1166 1167@item @emph{C/C++}: 1168@multitable @columnfractions .20 .80 1169@item @emph{Prototype}: @tab @code{void omp_set_nest_lock(omp_nest_lock_t *lock);} 1170@end multitable 1171 1172@item @emph{Fortran}: 1173@multitable @columnfractions .20 .80 1174@item @emph{Interface}: @tab @code{subroutine omp_set_nest_lock(nvar)} 1175@item @tab @code{integer(omp_nest_lock_kind), intent(inout) :: nvar} 1176@end multitable 1177 1178@item @emph{See also}: 1179@ref{omp_init_nest_lock}, @ref{omp_unset_nest_lock} 1180 1181@item @emph{Reference}: 1182@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.3.4. 1183@end table 1184 1185 1186 1187@node omp_test_nest_lock 1188@section @code{omp_test_nest_lock} -- Test and set nested lock if available 1189@table @asis 1190@item @emph{Description}: 1191Before setting a nested lock, the lock variable must be initialized by 1192@code{omp_init_nest_lock}. Contrary to @code{omp_set_nest_lock}, 1193@code{omp_test_nest_lock} does not block if the lock is not available. 1194If the lock is already held by the current thread, the new nesting count 1195is returned. Otherwise, the return value equals zero. 1196 1197@item @emph{C/C++}: 1198@multitable @columnfractions .20 .80 1199@item @emph{Prototype}: @tab @code{int omp_test_nest_lock(omp_nest_lock_t *lock);} 1200@end multitable 1201 1202@item @emph{Fortran}: 1203@multitable @columnfractions .20 .80 1204@item @emph{Interface}: @tab @code{logical function omp_test_nest_lock(nvar)} 1205@item @tab @code{integer(omp_nest_lock_kind), intent(inout) :: nvar} 1206@end multitable 1207 1208 1209@item @emph{See also}: 1210@ref{omp_init_lock}, @ref{omp_set_lock}, @ref{omp_set_lock} 1211 1212@item @emph{Reference}: 1213@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.3.6. 1214@end table 1215 1216 1217 1218@node omp_unset_nest_lock 1219@section @code{omp_unset_nest_lock} -- Unset nested lock 1220@table @asis 1221@item @emph{Description}: 1222A nested lock about to be unset must have been locked by @code{omp_set_nested_lock} 1223or @code{omp_test_nested_lock} before. In addition, the lock must be held by the 1224thread calling @code{omp_unset_nested_lock}. If the nesting count drops to zero, the 1225lock becomes unlocked. If one ore more threads attempted to set the lock before, 1226one of them is chosen to, again, set the lock to itself. 1227 1228@item @emph{C/C++}: 1229@multitable @columnfractions .20 .80 1230@item @emph{Prototype}: @tab @code{void omp_unset_nest_lock(omp_nest_lock_t *lock);} 1231@end multitable 1232 1233@item @emph{Fortran}: 1234@multitable @columnfractions .20 .80 1235@item @emph{Interface}: @tab @code{subroutine omp_unset_nest_lock(nvar)} 1236@item @tab @code{integer(omp_nest_lock_kind), intent(inout) :: nvar} 1237@end multitable 1238 1239@item @emph{See also}: 1240@ref{omp_set_nest_lock} 1241 1242@item @emph{Reference}: 1243@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.3.5. 1244@end table 1245 1246 1247 1248@node omp_destroy_nest_lock 1249@section @code{omp_destroy_nest_lock} -- Destroy nested lock 1250@table @asis 1251@item @emph{Description}: 1252Destroy a nested lock. In order to be destroyed, a nested lock must be 1253in the unlocked state and its nesting count must equal zero. 1254 1255@item @emph{C/C++}: 1256@multitable @columnfractions .20 .80 1257@item @emph{Prototype}: @tab @code{void omp_destroy_nest_lock(omp_nest_lock_t *);} 1258@end multitable 1259 1260@item @emph{Fortran}: 1261@multitable @columnfractions .20 .80 1262@item @emph{Interface}: @tab @code{subroutine omp_destroy_nest_lock(nvar)} 1263@item @tab @code{integer(omp_nest_lock_kind), intent(inout) :: nvar} 1264@end multitable 1265 1266@item @emph{See also}: 1267@ref{omp_init_lock} 1268 1269@item @emph{Reference}: 1270@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.3.3. 1271@end table 1272 1273 1274 1275@node omp_get_wtick 1276@section @code{omp_get_wtick} -- Get timer precision 1277@table @asis 1278@item @emph{Description}: 1279Gets the timer precision, i.e., the number of seconds between two 1280successive clock ticks. 1281 1282@item @emph{C/C++}: 1283@multitable @columnfractions .20 .80 1284@item @emph{Prototype}: @tab @code{double omp_get_wtick(void);} 1285@end multitable 1286 1287@item @emph{Fortran}: 1288@multitable @columnfractions .20 .80 1289@item @emph{Interface}: @tab @code{double precision function omp_get_wtick()} 1290@end multitable 1291 1292@item @emph{See also}: 1293@ref{omp_get_wtime} 1294 1295@item @emph{Reference}: 1296@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.4.2. 1297@end table 1298 1299 1300 1301@node omp_get_wtime 1302@section @code{omp_get_wtime} -- Elapsed wall clock time 1303@table @asis 1304@item @emph{Description}: 1305Elapsed wall clock time in seconds. The time is measured per thread, no 1306guarantee can be made that two distinct threads measure the same time. 1307Time is measured from some "time in the past", which is an arbitrary time 1308guaranteed not to change during the execution of the program. 1309 1310@item @emph{C/C++}: 1311@multitable @columnfractions .20 .80 1312@item @emph{Prototype}: @tab @code{double omp_get_wtime(void);} 1313@end multitable 1314 1315@item @emph{Fortran}: 1316@multitable @columnfractions .20 .80 1317@item @emph{Interface}: @tab @code{double precision function omp_get_wtime()} 1318@end multitable 1319 1320@item @emph{See also}: 1321@ref{omp_get_wtick} 1322 1323@item @emph{Reference}: 1324@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 3.4.1. 1325@end table 1326 1327 1328 1329@c --------------------------------------------------------------------- 1330@c Environment Variables 1331@c --------------------------------------------------------------------- 1332 1333@node Environment Variables 1334@chapter Environment Variables 1335 1336The environment variables which beginning with @env{OMP_} are defined by 1337section 4 of the OpenMP specification in version 4.5, while those 1338beginning with @env{GOMP_} are GNU extensions. 1339 1340@menu 1341* OMP_CANCELLATION:: Set whether cancellation is activated 1342* OMP_DISPLAY_ENV:: Show OpenMP version and environment variables 1343* OMP_DEFAULT_DEVICE:: Set the device used in target regions 1344* OMP_DYNAMIC:: Dynamic adjustment of threads 1345* OMP_MAX_ACTIVE_LEVELS:: Set the maximum number of nested parallel regions 1346* OMP_MAX_TASK_PRIORITY:: Set the maximum task priority value 1347* OMP_NESTED:: Nested parallel regions 1348* OMP_NUM_THREADS:: Specifies the number of threads to use 1349* OMP_PROC_BIND:: Whether theads may be moved between CPUs 1350* OMP_PLACES:: Specifies on which CPUs the theads should be placed 1351* OMP_STACKSIZE:: Set default thread stack size 1352* OMP_SCHEDULE:: How threads are scheduled 1353* OMP_THREAD_LIMIT:: Set the maximum number of threads 1354* OMP_WAIT_POLICY:: How waiting threads are handled 1355* GOMP_CPU_AFFINITY:: Bind threads to specific CPUs 1356* GOMP_DEBUG:: Enable debugging output 1357* GOMP_STACKSIZE:: Set default thread stack size 1358* GOMP_SPINCOUNT:: Set the busy-wait spin count 1359* GOMP_RTEMS_THREAD_POOLS:: Set the RTEMS specific thread pools 1360@end menu 1361 1362 1363@node OMP_CANCELLATION 1364@section @env{OMP_CANCELLATION} -- Set whether cancellation is activated 1365@cindex Environment Variable 1366@table @asis 1367@item @emph{Description}: 1368If set to @code{TRUE}, the cancellation is activated. If set to @code{FALSE} or 1369if unset, cancellation is disabled and the @code{cancel} construct is ignored. 1370 1371@item @emph{See also}: 1372@ref{omp_get_cancellation} 1373 1374@item @emph{Reference}: 1375@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.11 1376@end table 1377 1378 1379 1380@node OMP_DISPLAY_ENV 1381@section @env{OMP_DISPLAY_ENV} -- Show OpenMP version and environment variables 1382@cindex Environment Variable 1383@table @asis 1384@item @emph{Description}: 1385If set to @code{TRUE}, the OpenMP version number and the values 1386associated with the OpenMP environment variables are printed to @code{stderr}. 1387If set to @code{VERBOSE}, it additionally shows the value of the environment 1388variables which are GNU extensions. If undefined or set to @code{FALSE}, 1389this information will not be shown. 1390 1391 1392@item @emph{Reference}: 1393@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.12 1394@end table 1395 1396 1397 1398@node OMP_DEFAULT_DEVICE 1399@section @env{OMP_DEFAULT_DEVICE} -- Set the device used in target regions 1400@cindex Environment Variable 1401@table @asis 1402@item @emph{Description}: 1403Set to choose the device which is used in a @code{target} region, unless the 1404value is overridden by @code{omp_set_default_device} or by a @code{device} 1405clause. The value shall be the nonnegative device number. If no device with 1406the given device number exists, the code is executed on the host. If unset, 1407device number 0 will be used. 1408 1409 1410@item @emph{See also}: 1411@ref{omp_get_default_device}, @ref{omp_set_default_device}, 1412 1413@item @emph{Reference}: 1414@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.13 1415@end table 1416 1417 1418 1419@node OMP_DYNAMIC 1420@section @env{OMP_DYNAMIC} -- Dynamic adjustment of threads 1421@cindex Environment Variable 1422@table @asis 1423@item @emph{Description}: 1424Enable or disable the dynamic adjustment of the number of threads 1425within a team. The value of this environment variable shall be 1426@code{TRUE} or @code{FALSE}. If undefined, dynamic adjustment is 1427disabled by default. 1428 1429@item @emph{See also}: 1430@ref{omp_set_dynamic} 1431 1432@item @emph{Reference}: 1433@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.3 1434@end table 1435 1436 1437 1438@node OMP_MAX_ACTIVE_LEVELS 1439@section @env{OMP_MAX_ACTIVE_LEVELS} -- Set the maximum number of nested parallel regions 1440@cindex Environment Variable 1441@table @asis 1442@item @emph{Description}: 1443Specifies the initial value for the maximum number of nested parallel 1444regions. The value of this variable shall be a positive integer. 1445If undefined, the number of active levels is unlimited. 1446 1447@item @emph{See also}: 1448@ref{omp_set_max_active_levels} 1449 1450@item @emph{Reference}: 1451@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.9 1452@end table 1453 1454 1455 1456@node OMP_MAX_TASK_PRIORITY 1457@section @env{OMP_MAX_TASK_PRIORITY} -- Set the maximum priority 1458number that can be set for a task. 1459@cindex Environment Variable 1460@table @asis 1461@item @emph{Description}: 1462Specifies the initial value for the maximum priority value that can be 1463set for a task. The value of this variable shall be a non-negative 1464integer, and zero is allowed. If undefined, the default priority is 14650. 1466 1467@item @emph{See also}: 1468@ref{omp_get_max_task_priority} 1469 1470@item @emph{Reference}: 1471@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.14 1472@end table 1473 1474 1475 1476@node OMP_NESTED 1477@section @env{OMP_NESTED} -- Nested parallel regions 1478@cindex Environment Variable 1479@cindex Implementation specific setting 1480@table @asis 1481@item @emph{Description}: 1482Enable or disable nested parallel regions, i.e., whether team members 1483are allowed to create new teams. The value of this environment variable 1484shall be @code{TRUE} or @code{FALSE}. If undefined, nested parallel 1485regions are disabled by default. 1486 1487@item @emph{See also}: 1488@ref{omp_set_nested} 1489 1490@item @emph{Reference}: 1491@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.6 1492@end table 1493 1494 1495 1496@node OMP_NUM_THREADS 1497@section @env{OMP_NUM_THREADS} -- Specifies the number of threads to use 1498@cindex Environment Variable 1499@cindex Implementation specific setting 1500@table @asis 1501@item @emph{Description}: 1502Specifies the default number of threads to use in parallel regions. The 1503value of this variable shall be a comma-separated list of positive integers; 1504the value specified the number of threads to use for the corresponding nested 1505level. If undefined one thread per CPU is used. 1506 1507@item @emph{See also}: 1508@ref{omp_set_num_threads} 1509 1510@item @emph{Reference}: 1511@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.2 1512@end table 1513 1514 1515 1516@node OMP_PROC_BIND 1517@section @env{OMP_PROC_BIND} -- Whether theads may be moved between CPUs 1518@cindex Environment Variable 1519@table @asis 1520@item @emph{Description}: 1521Specifies whether threads may be moved between processors. If set to 1522@code{TRUE}, OpenMP theads should not be moved; if set to @code{FALSE} 1523they may be moved. Alternatively, a comma separated list with the 1524values @code{MASTER}, @code{CLOSE} and @code{SPREAD} can be used to specify 1525the thread affinity policy for the corresponding nesting level. With 1526@code{MASTER} the worker threads are in the same place partition as the 1527master thread. With @code{CLOSE} those are kept close to the master thread 1528in contiguous place partitions. And with @code{SPREAD} a sparse distribution 1529across the place partitions is used. 1530 1531When undefined, @env{OMP_PROC_BIND} defaults to @code{TRUE} when 1532@env{OMP_PLACES} or @env{GOMP_CPU_AFFINITY} is set and @code{FALSE} otherwise. 1533 1534@item @emph{See also}: 1535@ref{OMP_PLACES}, @ref{GOMP_CPU_AFFINITY}, @ref{omp_get_proc_bind} 1536 1537@item @emph{Reference}: 1538@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.4 1539@end table 1540 1541 1542 1543@node OMP_PLACES 1544@section @env{OMP_PLACES} -- Specifies on which CPUs the theads should be placed 1545@cindex Environment Variable 1546@table @asis 1547@item @emph{Description}: 1548The thread placement can be either specified using an abstract name or by an 1549explicit list of the places. The abstract names @code{threads}, @code{cores} 1550and @code{sockets} can be optionally followed by a positive number in 1551parentheses, which denotes the how many places shall be created. With 1552@code{threads} each place corresponds to a single hardware thread; @code{cores} 1553to a single core with the corresponding number of hardware threads; and with 1554@code{sockets} the place corresponds to a single socket. The resulting 1555placement can be shown by setting the @env{OMP_DISPLAY_ENV} environment 1556variable. 1557 1558Alternatively, the placement can be specified explicitly as comma-separated 1559list of places. A place is specified by set of nonnegative numbers in curly 1560braces, denoting the denoting the hardware threads. The hardware threads 1561belonging to a place can either be specified as comma-separated list of 1562nonnegative thread numbers or using an interval. Multiple places can also be 1563either specified by a comma-separated list of places or by an interval. To 1564specify an interval, a colon followed by the count is placed after after 1565the hardware thread number or the place. Optionally, the length can be 1566followed by a colon and the stride number -- otherwise a unit stride is 1567assumed. For instance, the following specifies the same places list: 1568@code{"@{0,1,2@}, @{3,4,6@}, @{7,8,9@}, @{10,11,12@}"}; 1569@code{"@{0:3@}, @{3:3@}, @{7:3@}, @{10:3@}"}; and @code{"@{0:2@}:4:3"}. 1570 1571If @env{OMP_PLACES} and @env{GOMP_CPU_AFFINITY} are unset and 1572@env{OMP_PROC_BIND} is either unset or @code{false}, threads may be moved 1573between CPUs following no placement policy. 1574 1575@item @emph{See also}: 1576@ref{OMP_PROC_BIND}, @ref{GOMP_CPU_AFFINITY}, @ref{omp_get_proc_bind}, 1577@ref{OMP_DISPLAY_ENV} 1578 1579@item @emph{Reference}: 1580@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.5 1581@end table 1582 1583 1584 1585@node OMP_STACKSIZE 1586@section @env{OMP_STACKSIZE} -- Set default thread stack size 1587@cindex Environment Variable 1588@table @asis 1589@item @emph{Description}: 1590Set the default thread stack size in kilobytes, unless the number 1591is suffixed by @code{B}, @code{K}, @code{M} or @code{G}, in which 1592case the size is, respectively, in bytes, kilobytes, megabytes 1593or gigabytes. This is different from @code{pthread_attr_setstacksize} 1594which gets the number of bytes as an argument. If the stack size cannot 1595be set due to system constraints, an error is reported and the initial 1596stack size is left unchanged. If undefined, the stack size is system 1597dependent. 1598 1599@item @emph{Reference}: 1600@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.7 1601@end table 1602 1603 1604 1605@node OMP_SCHEDULE 1606@section @env{OMP_SCHEDULE} -- How threads are scheduled 1607@cindex Environment Variable 1608@cindex Implementation specific setting 1609@table @asis 1610@item @emph{Description}: 1611Allows to specify @code{schedule type} and @code{chunk size}. 1612The value of the variable shall have the form: @code{type[,chunk]} where 1613@code{type} is one of @code{static}, @code{dynamic}, @code{guided} or @code{auto} 1614The optional @code{chunk} size shall be a positive integer. If undefined, 1615dynamic scheduling and a chunk size of 1 is used. 1616 1617@item @emph{See also}: 1618@ref{omp_set_schedule} 1619 1620@item @emph{Reference}: 1621@uref{http://www.openmp.org/, OpenMP specification v4.5}, Sections 2.7.1.1 and 4.1 1622@end table 1623 1624 1625 1626@node OMP_THREAD_LIMIT 1627@section @env{OMP_THREAD_LIMIT} -- Set the maximum number of threads 1628@cindex Environment Variable 1629@table @asis 1630@item @emph{Description}: 1631Specifies the number of threads to use for the whole program. The 1632value of this variable shall be a positive integer. If undefined, 1633the number of threads is not limited. 1634 1635@item @emph{See also}: 1636@ref{OMP_NUM_THREADS}, @ref{omp_get_thread_limit} 1637 1638@item @emph{Reference}: 1639@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.10 1640@end table 1641 1642 1643 1644@node OMP_WAIT_POLICY 1645@section @env{OMP_WAIT_POLICY} -- How waiting threads are handled 1646@cindex Environment Variable 1647@table @asis 1648@item @emph{Description}: 1649Specifies whether waiting threads should be active or passive. If 1650the value is @code{PASSIVE}, waiting threads should not consume CPU 1651power while waiting; while the value is @code{ACTIVE} specifies that 1652they should. If undefined, threads wait actively for a short time 1653before waiting passively. 1654 1655@item @emph{See also}: 1656@ref{GOMP_SPINCOUNT} 1657 1658@item @emph{Reference}: 1659@uref{http://www.openmp.org/, OpenMP specification v4.5}, Section 4.8 1660@end table 1661 1662 1663 1664@node GOMP_CPU_AFFINITY 1665@section @env{GOMP_CPU_AFFINITY} -- Bind threads to specific CPUs 1666@cindex Environment Variable 1667@table @asis 1668@item @emph{Description}: 1669Binds threads to specific CPUs. The variable should contain a space-separated 1670or comma-separated list of CPUs. This list may contain different kinds of 1671entries: either single CPU numbers in any order, a range of CPUs (M-N) 1672or a range with some stride (M-N:S). CPU numbers are zero based. For example, 1673@code{GOMP_CPU_AFFINITY="0 3 1-2 4-15:2"} will bind the initial thread 1674to CPU 0, the second to CPU 3, the third to CPU 1, the fourth to 1675CPU 2, the fifth to CPU 4, the sixth through tenth to CPUs 6, 8, 10, 12, 1676and 14 respectively and then start assigning back from the beginning of 1677the list. @code{GOMP_CPU_AFFINITY=0} binds all threads to CPU 0. 1678 1679There is no libgomp library routine to determine whether a CPU affinity 1680specification is in effect. As a workaround, language-specific library 1681functions, e.g., @code{getenv} in C or @code{GET_ENVIRONMENT_VARIABLE} in 1682Fortran, may be used to query the setting of the @code{GOMP_CPU_AFFINITY} 1683environment variable. A defined CPU affinity on startup cannot be changed 1684or disabled during the runtime of the application. 1685 1686If both @env{GOMP_CPU_AFFINITY} and @env{OMP_PROC_BIND} are set, 1687@env{OMP_PROC_BIND} has a higher precedence. If neither has been set and 1688@env{OMP_PROC_BIND} is unset, or when @env{OMP_PROC_BIND} is set to 1689@code{FALSE}, the host system will handle the assignment of threads to CPUs. 1690 1691@item @emph{See also}: 1692@ref{OMP_PLACES}, @ref{OMP_PROC_BIND} 1693@end table 1694 1695 1696 1697@node GOMP_DEBUG 1698@section @env{GOMP_DEBUG} -- Enable debugging output 1699@cindex Environment Variable 1700@table @asis 1701@item @emph{Description}: 1702Enable debugging output. The variable should be set to @code{0} 1703(disabled, also the default if not set), or @code{1} (enabled). 1704 1705If enabled, some debugging output will be printed during execution. 1706This is currently not specified in more detail, and subject to change. 1707@end table 1708 1709 1710 1711@node GOMP_STACKSIZE 1712@section @env{GOMP_STACKSIZE} -- Set default thread stack size 1713@cindex Environment Variable 1714@cindex Implementation specific setting 1715@table @asis 1716@item @emph{Description}: 1717Set the default thread stack size in kilobytes. This is different from 1718@code{pthread_attr_setstacksize} which gets the number of bytes as an 1719argument. If the stack size cannot be set due to system constraints, an 1720error is reported and the initial stack size is left unchanged. If undefined, 1721the stack size is system dependent. 1722 1723@item @emph{See also}: 1724@ref{OMP_STACKSIZE} 1725 1726@item @emph{Reference}: 1727@uref{http://gcc.gnu.org/ml/gcc-patches/2006-06/msg00493.html, 1728GCC Patches Mailinglist}, 1729@uref{http://gcc.gnu.org/ml/gcc-patches/2006-06/msg00496.html, 1730GCC Patches Mailinglist} 1731@end table 1732 1733 1734 1735@node GOMP_SPINCOUNT 1736@section @env{GOMP_SPINCOUNT} -- Set the busy-wait spin count 1737@cindex Environment Variable 1738@cindex Implementation specific setting 1739@table @asis 1740@item @emph{Description}: 1741Determines how long a threads waits actively with consuming CPU power 1742before waiting passively without consuming CPU power. The value may be 1743either @code{INFINITE}, @code{INFINITY} to always wait actively or an 1744integer which gives the number of spins of the busy-wait loop. The 1745integer may optionally be followed by the following suffixes acting 1746as multiplication factors: @code{k} (kilo, thousand), @code{M} (mega, 1747million), @code{G} (giga, billion), or @code{T} (tera, trillion). 1748If undefined, 0 is used when @env{OMP_WAIT_POLICY} is @code{PASSIVE}, 1749300,000 is used when @env{OMP_WAIT_POLICY} is undefined and 175030 billion is used when @env{OMP_WAIT_POLICY} is @code{ACTIVE}. 1751If there are more OpenMP threads than available CPUs, 1000 and 100 1752spins are used for @env{OMP_WAIT_POLICY} being @code{ACTIVE} or 1753undefined, respectively; unless the @env{GOMP_SPINCOUNT} is lower 1754or @env{OMP_WAIT_POLICY} is @code{PASSIVE}. 1755 1756@item @emph{See also}: 1757@ref{OMP_WAIT_POLICY} 1758@end table 1759 1760 1761 1762@node GOMP_RTEMS_THREAD_POOLS 1763@section @env{GOMP_RTEMS_THREAD_POOLS} -- Set the RTEMS specific thread pools 1764@cindex Environment Variable 1765@cindex Implementation specific setting 1766@table @asis 1767@item @emph{Description}: 1768This environment variable is only used on the RTEMS real-time operating system. 1769It determines the scheduler instance specific thread pools. The format for 1770@env{GOMP_RTEMS_THREAD_POOLS} is a list of optional 1771@code{<thread-pool-count>[$<priority>]@@<scheduler-name>} configurations 1772separated by @code{:} where: 1773@itemize @bullet 1774@item @code{<thread-pool-count>} is the thread pool count for this scheduler 1775instance. 1776@item @code{$<priority>} is an optional priority for the worker threads of a 1777thread pool according to @code{pthread_setschedparam}. In case a priority 1778value is omitted, then a worker thread will inherit the priority of the OpenMP 1779master thread that created it. The priority of the worker thread is not 1780changed after creation, even if a new OpenMP master thread using the worker has 1781a different priority. 1782@item @code{@@<scheduler-name>} is the scheduler instance name according to the 1783RTEMS application configuration. 1784@end itemize 1785In case no thread pool configuration is specified for a scheduler instance, 1786then each OpenMP master thread of this scheduler instance will use its own 1787dynamically allocated thread pool. To limit the worker thread count of the 1788thread pools, each OpenMP master thread must call @code{omp_set_num_threads}. 1789@item @emph{Example}: 1790Lets suppose we have three scheduler instances @code{IO}, @code{WRK0}, and 1791@code{WRK1} with @env{GOMP_RTEMS_THREAD_POOLS} set to 1792@code{"1@@WRK0:3$4@@WRK1"}. Then there are no thread pool restrictions for 1793scheduler instance @code{IO}. In the scheduler instance @code{WRK0} there is 1794one thread pool available. Since no priority is specified for this scheduler 1795instance, the worker thread inherits the priority of the OpenMP master thread 1796that created it. In the scheduler instance @code{WRK1} there are three thread 1797pools available and their worker threads run at priority four. 1798@end table 1799 1800 1801 1802@c --------------------------------------------------------------------- 1803@c Enabling OpenACC 1804@c --------------------------------------------------------------------- 1805 1806@node Enabling OpenACC 1807@chapter Enabling OpenACC 1808 1809To activate the OpenACC extensions for C/C++ and Fortran, the compile-time 1810flag @option{-fopenacc} must be specified. This enables the OpenACC directive 1811@code{#pragma acc} in C/C++ and @code{!$accp} directives in free form, 1812@code{c$acc}, @code{*$acc} and @code{!$acc} directives in fixed form, 1813@code{!$} conditional compilation sentinels in free form and @code{c$}, 1814@code{*$} and @code{!$} sentinels in fixed form, for Fortran. The flag also 1815arranges for automatic linking of the OpenACC runtime library 1816(@ref{OpenACC Runtime Library Routines}). 1817 1818A complete description of all OpenACC directives accepted may be found in 1819the @uref{http://www.openacc.org/, OpenACC} Application Programming 1820Interface manual, version 2.0. 1821 1822Note that this is an experimental feature and subject to 1823change in future versions of GCC. See 1824@uref{https://gcc.gnu.org/wiki/OpenACC} for more information. 1825 1826 1827 1828@c --------------------------------------------------------------------- 1829@c OpenACC Runtime Library Routines 1830@c --------------------------------------------------------------------- 1831 1832@node OpenACC Runtime Library Routines 1833@chapter OpenACC Runtime Library Routines 1834 1835The runtime routines described here are defined by section 3 of the OpenACC 1836specifications in version 2.0. 1837They have C linkage, and do not throw exceptions. 1838Generally, they are available only for the host, with the exception of 1839@code{acc_on_device}, which is available for both the host and the 1840acceleration device. 1841 1842@menu 1843* acc_get_num_devices:: Get number of devices for the given device 1844 type. 1845* acc_set_device_type:: Set type of device accelerator to use. 1846* acc_get_device_type:: Get type of device accelerator to be used. 1847* acc_set_device_num:: Set device number to use. 1848* acc_get_device_num:: Get device number to be used. 1849* acc_async_test:: Tests for completion of a specific asynchronous 1850 operation. 1851* acc_async_test_all:: Tests for completion of all asychronous 1852 operations. 1853* acc_wait:: Wait for completion of a specific asynchronous 1854 operation. 1855* acc_wait_all:: Waits for completion of all asyncrhonous 1856 operations. 1857* acc_wait_all_async:: Wait for completion of all asynchronous 1858 operations. 1859* acc_wait_async:: Wait for completion of asynchronous operations. 1860* acc_init:: Initialize runtime for a specific device type. 1861* acc_shutdown:: Shuts down the runtime for a specific device 1862 type. 1863* acc_on_device:: Whether executing on a particular device 1864* acc_malloc:: Allocate device memory. 1865* acc_free:: Free device memory. 1866* acc_copyin:: Allocate device memory and copy host memory to 1867 it. 1868* acc_present_or_copyin:: If the data is not present on the device, 1869 allocate device memory and copy from host 1870 memory. 1871* acc_create:: Allocate device memory and map it to host 1872 memory. 1873* acc_present_or_create:: If the data is not present on the device, 1874 allocate device memory and map it to host 1875 memory. 1876* acc_copyout:: Copy device memory to host memory. 1877* acc_delete:: Free device memory. 1878* acc_update_device:: Update device memory from mapped host memory. 1879* acc_update_self:: Update host memory from mapped device memory. 1880* acc_map_data:: Map previously allocated device memory to host 1881 memory. 1882* acc_unmap_data:: Unmap device memory from host memory. 1883* acc_deviceptr:: Get device pointer associated with specific 1884 host address. 1885* acc_hostptr:: Get host pointer associated with specific 1886 device address. 1887* acc_is_present:: Indiciate whether host variable / array is 1888 present on device. 1889* acc_memcpy_to_device:: Copy host memory to device memory. 1890* acc_memcpy_from_device:: Copy device memory to host memory. 1891 1892API routines for target platforms. 1893 1894* acc_get_current_cuda_device:: Get CUDA device handle. 1895* acc_get_current_cuda_context::Get CUDA context handle. 1896* acc_get_cuda_stream:: Get CUDA stream handle. 1897* acc_set_cuda_stream:: Set CUDA stream handle. 1898@end menu 1899 1900 1901 1902@node acc_get_num_devices 1903@section @code{acc_get_num_devices} -- Get number of devices for given device type 1904@table @asis 1905@item @emph{Description} 1906This function returns a value indicating the number of devices available 1907for the device type specified in @var{devicetype}. 1908 1909@item @emph{C/C++}: 1910@multitable @columnfractions .20 .80 1911@item @emph{Prototype}: @tab @code{int acc_get_num_devices(acc_device_t devicetype);} 1912@end multitable 1913 1914@item @emph{Fortran}: 1915@multitable @columnfractions .20 .80 1916@item @emph{Interface}: @tab @code{integer function acc_get_num_devices(devicetype)} 1917@item @tab @code{integer(kind=acc_device_kind) devicetype} 1918@end multitable 1919 1920@item @emph{Reference}: 1921@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 19223.2.1. 1923@end table 1924 1925 1926 1927@node acc_set_device_type 1928@section @code{acc_set_device_type} -- Set type of device accelerator to use. 1929@table @asis 1930@item @emph{Description} 1931This function indicates to the runtime library which device typr, specified 1932in @var{devicetype}, to use when executing a parallel or kernels region. 1933 1934@item @emph{C/C++}: 1935@multitable @columnfractions .20 .80 1936@item @emph{Prototype}: @tab @code{acc_set_device_type(acc_device_t devicetype);} 1937@end multitable 1938 1939@item @emph{Fortran}: 1940@multitable @columnfractions .20 .80 1941@item @emph{Interface}: @tab @code{subroutine acc_set_device_type(devicetype)} 1942@item @tab @code{integer(kind=acc_device_kind) devicetype} 1943@end multitable 1944 1945@item @emph{Reference}: 1946@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 19473.2.2. 1948@end table 1949 1950 1951 1952@node acc_get_device_type 1953@section @code{acc_get_device_type} -- Get type of device accelerator to be used. 1954@table @asis 1955@item @emph{Description} 1956This function returns what device type will be used when executing a 1957parallel or kernels region. 1958 1959@item @emph{C/C++}: 1960@multitable @columnfractions .20 .80 1961@item @emph{Prototype}: @tab @code{acc_device_t acc_get_device_type(void);} 1962@end multitable 1963 1964@item @emph{Fortran}: 1965@multitable @columnfractions .20 .80 1966@item @emph{Interface}: @tab @code{function acc_get_device_type(void)} 1967@item @tab @code{integer(kind=acc_device_kind) acc_get_device_type} 1968@end multitable 1969 1970@item @emph{Reference}: 1971@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 19723.2.3. 1973@end table 1974 1975 1976 1977@node acc_set_device_num 1978@section @code{acc_set_device_num} -- Set device number to use. 1979@table @asis 1980@item @emph{Description} 1981This function will indicate to the runtime which device number, 1982specified by @var{num}, associated with the specifed device 1983type @var{devicetype}. 1984 1985@item @emph{C/C++}: 1986@multitable @columnfractions .20 .80 1987@item @emph{Prototype}: @tab @code{acc_set_device_num(int num, acc_device_t devicetype);} 1988@end multitable 1989 1990@item @emph{Fortran}: 1991@multitable @columnfractions .20 .80 1992@item @emph{Interface}: @tab @code{subroutine acc_set_device_num(devicenum, devicetype)} 1993@item @tab @code{integer devicenum} 1994@item @tab @code{integer(kind=acc_device_kind) devicetype} 1995@end multitable 1996 1997@item @emph{Reference}: 1998@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 19993.2.4. 2000@end table 2001 2002 2003 2004@node acc_get_device_num 2005@section @code{acc_get_device_num} -- Get device number to be used. 2006@table @asis 2007@item @emph{Description} 2008This function returns which device number associated with the specified device 2009type @var{devicetype}, will be used when executing a parallel or kernels 2010region. 2011 2012@item @emph{C/C++}: 2013@multitable @columnfractions .20 .80 2014@item @emph{Prototype}: @tab @code{int acc_get_device_num(acc_device_t devicetype);} 2015@end multitable 2016 2017@item @emph{Fortran}: 2018@multitable @columnfractions .20 .80 2019@item @emph{Interface}: @tab @code{function acc_get_device_num(devicetype)} 2020@item @tab @code{integer(kind=acc_device_kind) devicetype} 2021@item @tab @code{integer acc_get_device_num} 2022@end multitable 2023 2024@item @emph{Reference}: 2025@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 20263.2.5. 2027@end table 2028 2029 2030 2031@node acc_async_test 2032@section @code{acc_async_test} -- Test for completion of a specific asynchronous operation. 2033@table @asis 2034@item @emph{Description} 2035This function tests for completion of the asynchrounous operation specified 2036in @var{arg}. In C/C++, a non-zero value will be returned to indicate 2037the specified asynchronous operation has completed. While Fortran will return 2038a @code{true}. If the asynchrounous operation has not completed, C/C++ returns 2039a zero and Fortran returns a @code{false}. 2040 2041@item @emph{C/C++}: 2042@multitable @columnfractions .20 .80 2043@item @emph{Prototype}: @tab @code{int acc_async_test(int arg);} 2044@end multitable 2045 2046@item @emph{Fortran}: 2047@multitable @columnfractions .20 .80 2048@item @emph{Interface}: @tab @code{function acc_async_test(arg)} 2049@item @tab @code{integer(kind=acc_handle_kind) arg} 2050@item @tab @code{logical acc_async_test} 2051@end multitable 2052 2053@item @emph{Reference}: 2054@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 20553.2.6. 2056@end table 2057 2058 2059 2060@node acc_async_test_all 2061@section @code{acc_async_test_all} -- Tests for completion of all asynchronous operations. 2062@table @asis 2063@item @emph{Description} 2064This function tests for completion of all asynchrounous operations. 2065In C/C++, a non-zero value will be returned to indicate all asynchronous 2066operations have completed. While Fortran will return a @code{true}. If 2067any asynchronous operation has not completed, C/C++ returns a zero and 2068Fortran returns a @code{false}. 2069 2070@item @emph{C/C++}: 2071@multitable @columnfractions .20 .80 2072@item @emph{Prototype}: @tab @code{int acc_async_test_all(void);} 2073@end multitable 2074 2075@item @emph{Fortran}: 2076@multitable @columnfractions .20 .80 2077@item @emph{Interface}: @tab @code{function acc_async_test()} 2078@item @tab @code{logical acc_get_device_num} 2079@end multitable 2080 2081@item @emph{Reference}: 2082@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 20833.2.7. 2084@end table 2085 2086 2087 2088@node acc_wait 2089@section @code{acc_wait} -- Wait for completion of a specific asynchronous operation. 2090@table @asis 2091@item @emph{Description} 2092This function waits for completion of the asynchronous operation 2093specified in @var{arg}. 2094 2095@item @emph{C/C++}: 2096@multitable @columnfractions .20 .80 2097@item @emph{Prototype}: @tab @code{acc_wait(arg);} 2098@end multitable 2099 2100@item @emph{Fortran}: 2101@multitable @columnfractions .20 .80 2102@item @emph{Interface}: @tab @code{subroutine acc_wait(arg)} 2103@item @tab @code{integer(acc_handle_kind) arg} 2104@end multitable 2105 2106@item @emph{Reference}: 2107@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 21083.2.8. 2109@end table 2110 2111 2112 2113@node acc_wait_all 2114@section @code{acc_wait_all} -- Waits for completion of all asynchronous operations. 2115@table @asis 2116@item @emph{Description} 2117This function waits for the completion of all asynchronous operations. 2118 2119@item @emph{C/C++}: 2120@multitable @columnfractions .20 .80 2121@item @emph{Prototype}: @tab @code{acc_wait_all(void);} 2122@end multitable 2123 2124@item @emph{Fortran}: 2125@multitable @columnfractions .20 .80 2126@item @emph{Interface}: @tab @code{subroutine acc_wait_async()} 2127@end multitable 2128 2129@item @emph{Reference}: 2130@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 21313.2.10. 2132@end table 2133 2134 2135 2136@node acc_wait_all_async 2137@section @code{acc_wait_all_async} -- Wait for completion of all asynchronous operations. 2138@table @asis 2139@item @emph{Description} 2140This function enqueues a wait operation on the queue @var{async} for any 2141and all asynchronous operations that have been previously enqueued on 2142any queue. 2143 2144@item @emph{C/C++}: 2145@multitable @columnfractions .20 .80 2146@item @emph{Prototype}: @tab @code{acc_wait_all_async(int async);} 2147@end multitable 2148 2149@item @emph{Fortran}: 2150@multitable @columnfractions .20 .80 2151@item @emph{Interface}: @tab @code{subroutine acc_wait_all_async(async)} 2152@item @tab @code{integer(acc_handle_kind) async} 2153@end multitable 2154 2155@item @emph{Reference}: 2156@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 21573.2.11. 2158@end table 2159 2160 2161 2162@node acc_wait_async 2163@section @code{acc_wait_async} -- Wait for completion of asynchronous operations. 2164@table @asis 2165@item @emph{Description} 2166This function enqueues a wait operation on queue @var{async} for any and all 2167asynchronous operations enqueued on queue @var{arg}. 2168 2169@item @emph{C/C++}: 2170@multitable @columnfractions .20 .80 2171@item @emph{Prototype}: @tab @code{acc_wait_async(int arg, int async);} 2172@end multitable 2173 2174@item @emph{Fortran}: 2175@multitable @columnfractions .20 .80 2176@item @emph{Interface}: @tab @code{subroutine acc_wait_async(arg, async)} 2177@item @tab @code{integer(acc_handle_kind) arg, async} 2178@end multitable 2179 2180@item @emph{Reference}: 2181@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 21823.2.9. 2183@end table 2184 2185 2186 2187@node acc_init 2188@section @code{acc_init} -- Initialize runtime for a specific device type. 2189@table @asis 2190@item @emph{Description} 2191This function initializes the runtime for the device type specified in 2192@var{devicetype}. 2193 2194@item @emph{C/C++}: 2195@multitable @columnfractions .20 .80 2196@item @emph{Prototype}: @tab @code{acc_init(acc_device_t devicetype);} 2197@end multitable 2198 2199@item @emph{Fortran}: 2200@multitable @columnfractions .20 .80 2201@item @emph{Interface}: @tab @code{subroutine acc_init(devicetype)} 2202@item @tab @code{integer(acc_device_kind) devicetype} 2203@end multitable 2204 2205@item @emph{Reference}: 2206@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 22073.2.12. 2208@end table 2209 2210 2211 2212@node acc_shutdown 2213@section @code{acc_shutdown} -- Shuts down the runtime for a specific device type. 2214@table @asis 2215@item @emph{Description} 2216This function shuts down the runtime for the device type specified in 2217@var{devicetype}. 2218 2219@item @emph{C/C++}: 2220@multitable @columnfractions .20 .80 2221@item @emph{Prototype}: @tab @code{acc_shutdown(acc_device_t devicetype);} 2222@end multitable 2223 2224@item @emph{Fortran}: 2225@multitable @columnfractions .20 .80 2226@item @emph{Interface}: @tab @code{subroutine acc_shutdown(devicetype)} 2227@item @tab @code{integer(acc_device_kind) devicetype} 2228@end multitable 2229 2230@item @emph{Reference}: 2231@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 22323.2.13. 2233@end table 2234 2235 2236 2237@node acc_on_device 2238@section @code{acc_on_device} -- Whether executing on a particular device 2239@table @asis 2240@item @emph{Description}: 2241This function returns whether the program is executing on a particular 2242device specified in @var{devicetype}. In C/C++ a non-zero value is 2243returned to indicate the device is execiting on the specified device type. 2244In Fortran, @code{true} will be returned. If the program is not executing 2245on the specified device type C/C++ will return a zero, while Fortran will 2246return @code{false}. 2247 2248@item @emph{C/C++}: 2249@multitable @columnfractions .20 .80 2250@item @emph{Prototype}: @tab @code{acc_on_device(acc_device_t devicetype);} 2251@end multitable 2252 2253@item @emph{Fortran}: 2254@multitable @columnfractions .20 .80 2255@item @emph{Interface}: @tab @code{function acc_on_device(devicetype)} 2256@item @tab @code{integer(acc_device_kind) devicetype} 2257@item @tab @code{logical acc_on_device} 2258@end multitable 2259 2260 2261@item @emph{Reference}: 2262@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 22633.2.14. 2264@end table 2265 2266 2267 2268@node acc_malloc 2269@section @code{acc_malloc} -- Allocate device memory. 2270@table @asis 2271@item @emph{Description} 2272This function allocates @var{len} bytes of device memory. It returns 2273the device address of the allocated memory. 2274 2275@item @emph{C/C++}: 2276@multitable @columnfractions .20 .80 2277@item @emph{Prototype}: @tab @code{d_void* acc_malloc(size_t len);} 2278@end multitable 2279 2280@item @emph{Reference}: 2281@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 22823.2.15. 2283@end table 2284 2285 2286 2287@node acc_free 2288@section @code{acc_free} -- Free device memory. 2289@table @asis 2290@item @emph{Description} 2291Free previously allocated device memory at the device address @code{a}. 2292 2293@item @emph{C/C++}: 2294@multitable @columnfractions .20 .80 2295@item @emph{Prototype}: @tab @code{acc_free(d_void *a);} 2296@end multitable 2297 2298@item @emph{Reference}: 2299@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 23003.2.16. 2301@end table 2302 2303 2304 2305@node acc_copyin 2306@section @code{acc_copyin} -- Allocate device memory and copy host memory to it. 2307@table @asis 2308@item @emph{Description} 2309In C/C++, this function allocates @var{len} bytes of device memory 2310and maps it to the specified host address in @var{a}. The device 2311address of the newly allocated device memory is returned. 2312 2313In Fortran, two (2) forms are supported. In the first form, @var{a} specifies 2314a contiguous array section. The second form @var{a} specifies a 2315variable or array element and @var{len} specifies the length in bytes. 2316 2317@item @emph{C/C++}: 2318@multitable @columnfractions .20 .80 2319@item @emph{Prototype}: @tab @code{void *acc_copyin(h_void *a, size_t len);} 2320@end multitable 2321 2322@item @emph{Fortran}: 2323@multitable @columnfractions .20 .80 2324@item @emph{Interface}: @tab @code{subroutine acc_copyin(a)} 2325@item @tab @code{type, dimension(:[,:]...) :: a} 2326@item @emph{Interface}: @tab @code{subroutine acc_copyin(a, len)} 2327@item @tab @code{type, dimension(:[,:]...) :: a} 2328@item @tab @code{integer len} 2329@end multitable 2330 2331@item @emph{Reference}: 2332@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 23333.2.17. 2334@end table 2335 2336 2337 2338@node acc_present_or_copyin 2339@section @code{acc_present_or_copyin} -- If the data is not present on the device, allocate device memory and copy from host memory. 2340@table @asis 2341@item @emph{Description} 2342This function tests if the host data specifed by @var{a} and of length 2343@var{len} is present or not. If it is not present, then device memory 2344will be allocated and the host memory copied. The device address of 2345the newly allocated device memory is returned. 2346 2347In Fortran, two (2) forms are supported. In the first form, @var{a} specifies 2348a contiguous array section. The second form @var{a} specifies a variable or 2349array element and @var{len} specifies the length in bytes. 2350 2351@item @emph{C/C++}: 2352@multitable @columnfractions .20 .80 2353@item @emph{Prototype}: @tab @code{void *acc_present_or_copyin(h_void *a, size_t len);} 2354@item @emph{Prototype}: @tab @code{void *acc_pcopyin(h_void *a, size_t len);} 2355@end multitable 2356 2357@item @emph{Fortran}: 2358@multitable @columnfractions .20 .80 2359@item @emph{Interface}: @tab @code{subroutine acc_present_or_copyin(a)} 2360@item @tab @code{type, dimension(:[,:]...) :: a} 2361@item @emph{Interface}: @tab @code{subroutine acc_present_or_copyin(a, len)} 2362@item @tab @code{type, dimension(:[,:]...) :: a} 2363@item @tab @code{integer len} 2364@item @emph{Interface}: @tab @code{subroutine acc_pcopyin(a)} 2365@item @tab @code{type, dimension(:[,:]...) :: a} 2366@item @emph{Interface}: @tab @code{subroutine acc_pcopyin(a, len)} 2367@item @tab @code{type, dimension(:[,:]...) :: a} 2368@item @tab @code{integer len} 2369@end multitable 2370 2371@item @emph{Reference}: 2372@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 23733.2.18. 2374@end table 2375 2376 2377 2378@node acc_create 2379@section @code{acc_create} -- Allocate device memory and map it to host memory. 2380@table @asis 2381@item @emph{Description} 2382This function allocates device memory and maps it to host memory specified 2383by the host address @var{a} with a length of @var{len} bytes. In C/C++, 2384the function returns the device address of the allocated device memory. 2385 2386In Fortran, two (2) forms are supported. In the first form, @var{a} specifies 2387a contiguous array section. The second form @var{a} specifies a variable or 2388array element and @var{len} specifies the length in bytes. 2389 2390@item @emph{C/C++}: 2391@multitable @columnfractions .20 .80 2392@item @emph{Prototype}: @tab @code{void *acc_create(h_void *a, size_t len);} 2393@end multitable 2394 2395@item @emph{Fortran}: 2396@multitable @columnfractions .20 .80 2397@item @emph{Interface}: @tab @code{subroutine acc_create(a)} 2398@item @tab @code{type, dimension(:[,:]...) :: a} 2399@item @emph{Interface}: @tab @code{subroutine acc_create(a, len)} 2400@item @tab @code{type, dimension(:[,:]...) :: a} 2401@item @tab @code{integer len} 2402@end multitable 2403 2404@item @emph{Reference}: 2405@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 24063.2.19. 2407@end table 2408 2409 2410 2411@node acc_present_or_create 2412@section @code{acc_present_or_create} -- If the data is not present on the device, allocate device memory and map it to host memory. 2413@table @asis 2414@item @emph{Description} 2415This function tests if the host data specifed by @var{a} and of length 2416@var{len} is present or not. If it is not present, then device memory 2417will be allocated and mapped to host memory. In C/C++, the device address 2418of the newly allocated device memory is returned. 2419 2420In Fortran, two (2) forms are supported. In the first form, @var{a} specifies 2421a contiguous array section. The second form @var{a} specifies a variable or 2422array element and @var{len} specifies the length in bytes. 2423 2424 2425@item @emph{C/C++}: 2426@multitable @columnfractions .20 .80 2427@item @emph{Prototype}: @tab @code{void *acc_present_or_create(h_void *a, size_t len)} 2428@item @emph{Prototype}: @tab @code{void *acc_pcreate(h_void *a, size_t len)} 2429@end multitable 2430 2431@item @emph{Fortran}: 2432@multitable @columnfractions .20 .80 2433@item @emph{Interface}: @tab @code{subroutine acc_present_or_create(a)} 2434@item @tab @code{type, dimension(:[,:]...) :: a} 2435@item @emph{Interface}: @tab @code{subroutine acc_present_or_create(a, len)} 2436@item @tab @code{type, dimension(:[,:]...) :: a} 2437@item @tab @code{integer len} 2438@item @emph{Interface}: @tab @code{subroutine acc_pcreate(a)} 2439@item @tab @code{type, dimension(:[,:]...) :: a} 2440@item @emph{Interface}: @tab @code{subroutine acc_pcreate(a, len)} 2441@item @tab @code{type, dimension(:[,:]...) :: a} 2442@item @tab @code{integer len} 2443@end multitable 2444 2445@item @emph{Reference}: 2446@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 24473.2.20. 2448@end table 2449 2450 2451 2452@node acc_copyout 2453@section @code{acc_copyout} -- Copy device memory to host memory. 2454@table @asis 2455@item @emph{Description} 2456This function copies mapped device memory to host memory which is specified 2457by host address @var{a} for a length @var{len} bytes in C/C++. 2458 2459In Fortran, two (2) forms are supported. In the first form, @var{a} specifies 2460a contiguous array section. The second form @var{a} specifies a variable or 2461array element and @var{len} specifies the length in bytes. 2462 2463@item @emph{C/C++}: 2464@multitable @columnfractions .20 .80 2465@item @emph{Prototype}: @tab @code{acc_copyout(h_void *a, size_t len);} 2466@end multitable 2467 2468@item @emph{Fortran}: 2469@multitable @columnfractions .20 .80 2470@item @emph{Interface}: @tab @code{subroutine acc_copyout(a)} 2471@item @tab @code{type, dimension(:[,:]...) :: a} 2472@item @emph{Interface}: @tab @code{subroutine acc_copyout(a, len)} 2473@item @tab @code{type, dimension(:[,:]...) :: a} 2474@item @tab @code{integer len} 2475@end multitable 2476 2477@item @emph{Reference}: 2478@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 24793.2.21. 2480@end table 2481 2482 2483 2484@node acc_delete 2485@section @code{acc_delete} -- Free device memory. 2486@table @asis 2487@item @emph{Description} 2488This function frees previously allocated device memory specified by 2489the device address @var{a} and the length of @var{len} bytes. 2490 2491In Fortran, two (2) forms are supported. In the first form, @var{a} specifies 2492a contiguous array section. The second form @var{a} specifies a variable or 2493array element and @var{len} specifies the length in bytes. 2494 2495@item @emph{C/C++}: 2496@multitable @columnfractions .20 .80 2497@item @emph{Prototype}: @tab @code{acc_delete(h_void *a, size_t len);} 2498@end multitable 2499 2500@item @emph{Fortran}: 2501@multitable @columnfractions .20 .80 2502@item @emph{Interface}: @tab @code{subroutine acc_delete(a)} 2503@item @tab @code{type, dimension(:[,:]...) :: a} 2504@item @emph{Interface}: @tab @code{subroutine acc_delete(a, len)} 2505@item @tab @code{type, dimension(:[,:]...) :: a} 2506@item @tab @code{integer len} 2507@end multitable 2508 2509@item @emph{Reference}: 2510@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 25113.2.22. 2512@end table 2513 2514 2515 2516@node acc_update_device 2517@section @code{acc_update_device} -- Update device memory from mapped host memory. 2518@table @asis 2519@item @emph{Description} 2520This function updates the device copy from the previously mapped host memory. 2521The host memory is specified with the host address @var{a} and a length of 2522@var{len} bytes. 2523 2524In Fortran, two (2) forms are supported. In the first form, @var{a} specifies 2525a contiguous array section. The second form @var{a} specifies a variable or 2526array element and @var{len} specifies the length in bytes. 2527 2528@item @emph{C/C++}: 2529@multitable @columnfractions .20 .80 2530@item @emph{Prototype}: @tab @code{acc_update_device(h_void *a, size_t len);} 2531@end multitable 2532 2533@item @emph{Fortran}: 2534@multitable @columnfractions .20 .80 2535@item @emph{Interface}: @tab @code{subroutine acc_update_device(a)} 2536@item @tab @code{type, dimension(:[,:]...) :: a} 2537@item @emph{Interface}: @tab @code{subroutine acc_update_device(a, len)} 2538@item @tab @code{type, dimension(:[,:]...) :: a} 2539@item @tab @code{integer len} 2540@end multitable 2541 2542@item @emph{Reference}: 2543@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 25443.2.23. 2545@end table 2546 2547 2548 2549@node acc_update_self 2550@section @code{acc_update_self} -- Update host memory from mapped device memory. 2551@table @asis 2552@item @emph{Description} 2553This function updates the host copy from the previously mapped device memory. 2554The host memory is specified with the host address @var{a} and a length of 2555@var{len} bytes. 2556 2557In Fortran, two (2) forms are supported. In the first form, @var{a} specifies 2558a contiguous array section. The second form @var{a} specifies a variable or 2559array element and @var{len} specifies the length in bytes. 2560 2561@item @emph{C/C++}: 2562@multitable @columnfractions .20 .80 2563@item @emph{Prototype}: @tab @code{acc_update_self(h_void *a, size_t len);} 2564@end multitable 2565 2566@item @emph{Fortran}: 2567@multitable @columnfractions .20 .80 2568@item @emph{Interface}: @tab @code{subroutine acc_update_self(a)} 2569@item @tab @code{type, dimension(:[,:]...) :: a} 2570@item @emph{Interface}: @tab @code{subroutine acc_update_self(a, len)} 2571@item @tab @code{type, dimension(:[,:]...) :: a} 2572@item @tab @code{integer len} 2573@end multitable 2574 2575@item @emph{Reference}: 2576@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 25773.2.24. 2578@end table 2579 2580 2581 2582@node acc_map_data 2583@section @code{acc_map_data} -- Map previously allocated device memory to host memory. 2584@table @asis 2585@item @emph{Description} 2586This function maps previously allocated device and host memory. The device 2587memory is specified with the device address @var{d}. The host memory is 2588specified with the host address @var{h} and a length of @var{len}. 2589 2590@item @emph{C/C++}: 2591@multitable @columnfractions .20 .80 2592@item @emph{Prototype}: @tab @code{acc_map_data(h_void *h, d_void *d, size_t len);} 2593@end multitable 2594 2595@item @emph{Reference}: 2596@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 25973.2.25. 2598@end table 2599 2600 2601 2602@node acc_unmap_data 2603@section @code{acc_unmap_data} -- Unmap device memory from host memory. 2604@table @asis 2605@item @emph{Description} 2606This function unmaps previously mapped device and host memory. The latter 2607specified by @var{h}. 2608 2609@item @emph{C/C++}: 2610@multitable @columnfractions .20 .80 2611@item @emph{Prototype}: @tab @code{acc_unmap_data(h_void *h);} 2612@end multitable 2613 2614@item @emph{Reference}: 2615@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 26163.2.26. 2617@end table 2618 2619 2620 2621@node acc_deviceptr 2622@section @code{acc_deviceptr} -- Get device pointer associated with specific host address. 2623@table @asis 2624@item @emph{Description} 2625This function returns the device address that has been mapped to the 2626host address specified by @var{h}. 2627 2628@item @emph{C/C++}: 2629@multitable @columnfractions .20 .80 2630@item @emph{Prototype}: @tab @code{void *acc_deviceptr(h_void *h);} 2631@end multitable 2632 2633@item @emph{Reference}: 2634@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 26353.2.27. 2636@end table 2637 2638 2639 2640@node acc_hostptr 2641@section @code{acc_hostptr} -- Get host pointer associated with specific device address. 2642@table @asis 2643@item @emph{Description} 2644This function returns the host address that has been mapped to the 2645device address specified by @var{d}. 2646 2647@item @emph{C/C++}: 2648@multitable @columnfractions .20 .80 2649@item @emph{Prototype}: @tab @code{void *acc_hostptr(d_void *d);} 2650@end multitable 2651 2652@item @emph{Reference}: 2653@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 26543.2.28. 2655@end table 2656 2657 2658 2659@node acc_is_present 2660@section @code{acc_is_present} -- Indicate whether host variable / array is present on device. 2661@table @asis 2662@item @emph{Description} 2663This function indicates whether the specified host address in @var{a} and a 2664length of @var{len} bytes is present on the device. In C/C++, a non-zero 2665value is returned to indicate the presence of the mapped memory on the 2666device. A zero is returned to indicate the memory is not mapped on the 2667device. 2668 2669In Fortran, two (2) forms are supported. In the first form, @var{a} specifies 2670a contiguous array section. The second form @var{a} specifies a variable or 2671array element and @var{len} specifies the length in bytes. If the host 2672memory is mapped to device memory, then a @code{true} is returned. Otherwise, 2673a @code{false} is return to indicate the mapped memory is not present. 2674 2675@item @emph{C/C++}: 2676@multitable @columnfractions .20 .80 2677@item @emph{Prototype}: @tab @code{int acc_is_present(h_void *a, size_t len);} 2678@end multitable 2679 2680@item @emph{Fortran}: 2681@multitable @columnfractions .20 .80 2682@item @emph{Interface}: @tab @code{function acc_is_present(a)} 2683@item @tab @code{type, dimension(:[,:]...) :: a} 2684@item @tab @code{logical acc_is_present} 2685@item @emph{Interface}: @tab @code{function acc_is_present(a, len)} 2686@item @tab @code{type, dimension(:[,:]...) :: a} 2687@item @tab @code{integer len} 2688@item @tab @code{logical acc_is_present} 2689@end multitable 2690 2691@item @emph{Reference}: 2692@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 26933.2.29. 2694@end table 2695 2696 2697 2698@node acc_memcpy_to_device 2699@section @code{acc_memcpy_to_device} -- Copy host memory to device memory. 2700@table @asis 2701@item @emph{Description} 2702This function copies host memory specified by host address of @var{src} to 2703device memory specified by the device address @var{dest} for a length of 2704@var{bytes} bytes. 2705 2706@item @emph{C/C++}: 2707@multitable @columnfractions .20 .80 2708@item @emph{Prototype}: @tab @code{acc_memcpy_to_device(d_void *dest, h_void *src, size_t bytes);} 2709@end multitable 2710 2711@item @emph{Reference}: 2712@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 27133.2.30. 2714@end table 2715 2716 2717 2718@node acc_memcpy_from_device 2719@section @code{acc_memcpy_from_device} -- Copy device memory to host memory. 2720@table @asis 2721@item @emph{Description} 2722This function copies host memory specified by host address of @var{src} from 2723device memory specified by the device address @var{dest} for a length of 2724@var{bytes} bytes. 2725 2726@item @emph{C/C++}: 2727@multitable @columnfractions .20 .80 2728@item @emph{Prototype}: @tab @code{acc_memcpy_from_device(d_void *dest, h_void *src, size_t bytes);} 2729@end multitable 2730 2731@item @emph{Reference}: 2732@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 27333.2.31. 2734@end table 2735 2736 2737 2738@node acc_get_current_cuda_device 2739@section @code{acc_get_current_cuda_device} -- Get CUDA device handle. 2740@table @asis 2741@item @emph{Description} 2742This function returns the CUDA device handle. This handle is the same 2743as used by the CUDA Runtime or Driver API's. 2744 2745@item @emph{C/C++}: 2746@multitable @columnfractions .20 .80 2747@item @emph{Prototype}: @tab @code{void *acc_get_current_cuda_device(void);} 2748@end multitable 2749 2750@item @emph{Reference}: 2751@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 2752A.2.1.1. 2753@end table 2754 2755 2756 2757@node acc_get_current_cuda_context 2758@section @code{acc_get_current_cuda_context} -- Get CUDA context handle. 2759@table @asis 2760@item @emph{Description} 2761This function returns the CUDA context handle. This handle is the same 2762as used by the CUDA Runtime or Driver API's. 2763 2764@item @emph{C/C++}: 2765@multitable @columnfractions .20 .80 2766@item @emph{Prototype}: @tab @code{acc_get_current_cuda_context(void);} 2767@end multitable 2768 2769@item @emph{Reference}: 2770@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 2771A.2.1.2. 2772@end table 2773 2774 2775 2776@node acc_get_cuda_stream 2777@section @code{acc_get_cuda_stream} -- Get CUDA stream handle. 2778@table @asis 2779@item @emph{Description} 2780This function returns the CUDA stream handle. This handle is the same 2781as used by the CUDA Runtime or Driver API's. 2782 2783@item @emph{C/C++}: 2784@multitable @columnfractions .20 .80 2785@item @emph{Prototype}: @tab @code{acc_get_cuda_stream(void);} 2786@end multitable 2787 2788@item @emph{Reference}: 2789@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 2790A.2.1.3. 2791@end table 2792 2793 2794 2795@node acc_set_cuda_stream 2796@section @code{acc_set_cuda_stream} -- Set CUDA stream handle. 2797@table @asis 2798@item @emph{Description} 2799This function associates the stream handle specified by @var{stream} with 2800the asynchronous value specified by @var{async}. 2801 2802@item @emph{C/C++}: 2803@multitable @columnfractions .20 .80 2804@item @emph{Prototype}: @tab @code{acc_set_cuda_stream(int async void *stream);} 2805@end multitable 2806 2807@item @emph{Reference}: 2808@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 2809A.2.1.4. 2810@end table 2811 2812 2813 2814@c --------------------------------------------------------------------- 2815@c OpenACC Environment Variables 2816@c --------------------------------------------------------------------- 2817 2818@node OpenACC Environment Variables 2819@chapter OpenACC Environment Variables 2820 2821The variables @env{ACC_DEVICE_TYPE} and @env{ACC_DEVICE_NUM} 2822are defined by section 4 of the OpenACC specification in version 2.0. 2823The variable @env{GCC_ACC_NOTIFY} is used for diagnostic purposes. 2824 2825@menu 2826* ACC_DEVICE_TYPE:: 2827* ACC_DEVICE_NUM:: 2828* GCC_ACC_NOTIFY:: 2829@end menu 2830 2831 2832 2833@node ACC_DEVICE_TYPE 2834@section @code{ACC_DEVICE_TYPE} 2835@table @asis 2836@item @emph{Reference}: 2837@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 28384.1. 2839@end table 2840 2841 2842 2843@node ACC_DEVICE_NUM 2844@section @code{ACC_DEVICE_NUM} 2845@table @asis 2846@item @emph{Reference}: 2847@uref{http://www.openacc.org/, OpenACC specification v2.0}, section 28484.2. 2849@end table 2850 2851 2852 2853@node GCC_ACC_NOTIFY 2854@section @code{GCC_ACC_NOTIFY} 2855@table @asis 2856@item @emph{Description}: 2857Print debug information pertaining to the accelerator. 2858@end table 2859 2860 2861 2862@c --------------------------------------------------------------------- 2863@c CUDA Streams Usage 2864@c --------------------------------------------------------------------- 2865 2866@node CUDA Streams Usage 2867@chapter CUDA Streams Usage 2868 2869This applies to the @code{nvptx} plugin only. 2870 2871The library provides elements that perform asynchronous movement of 2872data and asynchronous operation of computing constructs. This 2873asynchronous functionality is implemented by making use of CUDA 2874streams@footnote{See "Stream Management" in "CUDA Driver API", 2875TRM-06703-001, Version 5.5, for additional information}. 2876 2877The primary means by that the asychronous functionality is accessed 2878is through the use of those OpenACC directives which make use of the 2879@code{async} and @code{wait} clauses. When the @code{async} clause is 2880first used with a directive, it creates a CUDA stream. If an 2881@code{async-argument} is used with the @code{async} clause, then the 2882stream is associated with the specified @code{async-argument}. 2883 2884Following the creation of an association between a CUDA stream and the 2885@code{async-argument} of an @code{async} clause, both the @code{wait} 2886clause and the @code{wait} directive can be used. When either the 2887clause or directive is used after stream creation, it creates a 2888rendezvous point whereby execution waits until all operations 2889associated with the @code{async-argument}, that is, stream, have 2890completed. 2891 2892Normally, the management of the streams that are created as a result of 2893using the @code{async} clause, is done without any intervention by the 2894caller. This implies the association between the @code{async-argument} 2895and the CUDA stream will be maintained for the lifetime of the program. 2896However, this association can be changed through the use of the library 2897function @code{acc_set_cuda_stream}. When the function 2898@code{acc_set_cuda_stream} is called, the CUDA stream that was 2899originally associated with the @code{async} clause will be destroyed. 2900Caution should be taken when changing the association as subsequent 2901references to the @code{async-argument} refer to a different 2902CUDA stream. 2903 2904 2905 2906@c --------------------------------------------------------------------- 2907@c OpenACC Library Interoperability 2908@c --------------------------------------------------------------------- 2909 2910@node OpenACC Library Interoperability 2911@chapter OpenACC Library Interoperability 2912 2913@section Introduction 2914 2915The OpenACC library uses the CUDA Driver API, and may interact with 2916programs that use the Runtime library directly, or another library 2917based on the Runtime library, e.g., CUBLAS@footnote{See section 2.26, 2918"Interactions with the CUDA Driver API" in 2919"CUDA Runtime API", Version 5.5, and section 2.27, "VDPAU 2920Interoperability", in "CUDA Driver API", TRM-06703-001, Version 5.5, 2921for additional information on library interoperability.}. 2922This chapter describes the use cases and what changes are 2923required in order to use both the OpenACC library and the CUBLAS and Runtime 2924libraries within a program. 2925 2926@section First invocation: NVIDIA CUBLAS library API 2927 2928In this first use case (see below), a function in the CUBLAS library is called 2929prior to any of the functions in the OpenACC library. More specifically, the 2930function @code{cublasCreate()}. 2931 2932When invoked, the function initializes the library and allocates the 2933hardware resources on the host and the device on behalf of the caller. Once 2934the initialization and allocation has completed, a handle is returned to the 2935caller. The OpenACC library also requires initialization and allocation of 2936hardware resources. Since the CUBLAS library has already allocated the 2937hardware resources for the device, all that is left to do is to initialize 2938the OpenACC library and acquire the hardware resources on the host. 2939 2940Prior to calling the OpenACC function that initializes the library and 2941allocate the host hardware resources, you need to acquire the device number 2942that was allocated during the call to @code{cublasCreate()}. The invoking of the 2943runtime library function @code{cudaGetDevice()} accomplishes this. Once 2944acquired, the device number is passed along with the device type as 2945parameters to the OpenACC library function @code{acc_set_device_num()}. 2946 2947Once the call to @code{acc_set_device_num()} has completed, the OpenACC 2948library uses the context that was created during the call to 2949@code{cublasCreate()}. In other words, both libraries will be sharing the 2950same context. 2951 2952@smallexample 2953 /* Create the handle */ 2954 s = cublasCreate(&h); 2955 if (s != CUBLAS_STATUS_SUCCESS) 2956 @{ 2957 fprintf(stderr, "cublasCreate failed %d\n", s); 2958 exit(EXIT_FAILURE); 2959 @} 2960 2961 /* Get the device number */ 2962 e = cudaGetDevice(&dev); 2963 if (e != cudaSuccess) 2964 @{ 2965 fprintf(stderr, "cudaGetDevice failed %d\n", e); 2966 exit(EXIT_FAILURE); 2967 @} 2968 2969 /* Initialize OpenACC library and use device 'dev' */ 2970 acc_set_device_num(dev, acc_device_nvidia); 2971 2972@end smallexample 2973@center Use Case 1 2974 2975@section First invocation: OpenACC library API 2976 2977In this second use case (see below), a function in the OpenACC library is 2978called prior to any of the functions in the CUBLAS library. More specificially, 2979the function @code{acc_set_device_num()}. 2980 2981In the use case presented here, the function @code{acc_set_device_num()} 2982is used to both initialize the OpenACC library and allocate the hardware 2983resources on the host and the device. In the call to the function, the 2984call parameters specify which device to use and what device 2985type to use, i.e., @code{acc_device_nvidia}. It should be noted that this 2986is but one method to initialize the OpenACC library and allocate the 2987appropriate hardware resources. Other methods are available through the 2988use of environment variables and these will be discussed in the next section. 2989 2990Once the call to @code{acc_set_device_num()} has completed, other OpenACC 2991functions can be called as seen with multiple calls being made to 2992@code{acc_copyin()}. In addition, calls can be made to functions in the 2993CUBLAS library. In the use case a call to @code{cublasCreate()} is made 2994subsequent to the calls to @code{acc_copyin()}. 2995As seen in the previous use case, a call to @code{cublasCreate()} 2996initializes the CUBLAS library and allocates the hardware resources on the 2997host and the device. However, since the device has already been allocated, 2998@code{cublasCreate()} will only initialize the CUBLAS library and allocate 2999the appropriate hardware resources on the host. The context that was created 3000as part of the OpenACC initialization is shared with the CUBLAS library, 3001similarly to the first use case. 3002 3003@smallexample 3004 dev = 0; 3005 3006 acc_set_device_num(dev, acc_device_nvidia); 3007 3008 /* Copy the first set to the device */ 3009 d_X = acc_copyin(&h_X[0], N * sizeof (float)); 3010 if (d_X == NULL) 3011 @{ 3012 fprintf(stderr, "copyin error h_X\n"); 3013 exit(EXIT_FAILURE); 3014 @} 3015 3016 /* Copy the second set to the device */ 3017 d_Y = acc_copyin(&h_Y1[0], N * sizeof (float)); 3018 if (d_Y == NULL) 3019 @{ 3020 fprintf(stderr, "copyin error h_Y1\n"); 3021 exit(EXIT_FAILURE); 3022 @} 3023 3024 /* Create the handle */ 3025 s = cublasCreate(&h); 3026 if (s != CUBLAS_STATUS_SUCCESS) 3027 @{ 3028 fprintf(stderr, "cublasCreate failed %d\n", s); 3029 exit(EXIT_FAILURE); 3030 @} 3031 3032 /* Perform saxpy using CUBLAS library function */ 3033 s = cublasSaxpy(h, N, &alpha, d_X, 1, d_Y, 1); 3034 if (s != CUBLAS_STATUS_SUCCESS) 3035 @{ 3036 fprintf(stderr, "cublasSaxpy failed %d\n", s); 3037 exit(EXIT_FAILURE); 3038 @} 3039 3040 /* Copy the results from the device */ 3041 acc_memcpy_from_device(&h_Y1[0], d_Y, N * sizeof (float)); 3042 3043@end smallexample 3044@center Use Case 2 3045 3046@section OpenACC library and environment variables 3047 3048There are two environment variables associated with the OpenACC library 3049that may be used to control the device type and device number: 3050@env{ACC_DEVICE_TYPE} and @env{ACC_DEVICE_NUM}, respecively. These two 3051environement variables can be used as an alternative to calling 3052@code{acc_set_device_num()}. As seen in the second use case, the device 3053type and device number were specified using @code{acc_set_device_num()}. 3054If however, the aforementioned environment variables were set, then the 3055call to @code{acc_set_device_num()} would not be required. 3056 3057 3058The use of the environment variables is only relevant when an OpenACC function 3059is called prior to a call to @code{cudaCreate()}. If @code{cudaCreate()} 3060is called prior to a call to an OpenACC function, then you must call 3061@code{acc_set_device_num()}@footnote{More complete information 3062about @env{ACC_DEVICE_TYPE} and @env{ACC_DEVICE_NUM} can be found in 3063sections 4.1 and 4.2 of the @uref{http://www.openacc.org/, OpenACC} 3064Application Programming Interface”, Version 2.0.} 3065 3066 3067 3068@c --------------------------------------------------------------------- 3069@c The libgomp ABI 3070@c --------------------------------------------------------------------- 3071 3072@node The libgomp ABI 3073@chapter The libgomp ABI 3074 3075The following sections present notes on the external ABI as 3076presented by libgomp. Only maintainers should need them. 3077 3078@menu 3079* Implementing MASTER construct:: 3080* Implementing CRITICAL construct:: 3081* Implementing ATOMIC construct:: 3082* Implementing FLUSH construct:: 3083* Implementing BARRIER construct:: 3084* Implementing THREADPRIVATE construct:: 3085* Implementing PRIVATE clause:: 3086* Implementing FIRSTPRIVATE LASTPRIVATE COPYIN and COPYPRIVATE clauses:: 3087* Implementing REDUCTION clause:: 3088* Implementing PARALLEL construct:: 3089* Implementing FOR construct:: 3090* Implementing ORDERED construct:: 3091* Implementing SECTIONS construct:: 3092* Implementing SINGLE construct:: 3093* Implementing OpenACC's PARALLEL construct:: 3094@end menu 3095 3096 3097@node Implementing MASTER construct 3098@section Implementing MASTER construct 3099 3100@smallexample 3101if (omp_get_thread_num () == 0) 3102 block 3103@end smallexample 3104 3105Alternately, we generate two copies of the parallel subfunction 3106and only include this in the version run by the master thread. 3107Surely this is not worthwhile though... 3108 3109 3110 3111@node Implementing CRITICAL construct 3112@section Implementing CRITICAL construct 3113 3114Without a specified name, 3115 3116@smallexample 3117 void GOMP_critical_start (void); 3118 void GOMP_critical_end (void); 3119@end smallexample 3120 3121so that we don't get COPY relocations from libgomp to the main 3122application. 3123 3124With a specified name, use omp_set_lock and omp_unset_lock with 3125name being transformed into a variable declared like 3126 3127@smallexample 3128 omp_lock_t gomp_critical_user_<name> __attribute__((common)) 3129@end smallexample 3130 3131Ideally the ABI would specify that all zero is a valid unlocked 3132state, and so we wouldn't need to initialize this at 3133startup. 3134 3135 3136 3137@node Implementing ATOMIC construct 3138@section Implementing ATOMIC construct 3139 3140The target should implement the @code{__sync} builtins. 3141 3142Failing that we could add 3143 3144@smallexample 3145 void GOMP_atomic_enter (void) 3146 void GOMP_atomic_exit (void) 3147@end smallexample 3148 3149which reuses the regular lock code, but with yet another lock 3150object private to the library. 3151 3152 3153 3154@node Implementing FLUSH construct 3155@section Implementing FLUSH construct 3156 3157Expands to the @code{__sync_synchronize} builtin. 3158 3159 3160 3161@node Implementing BARRIER construct 3162@section Implementing BARRIER construct 3163 3164@smallexample 3165 void GOMP_barrier (void) 3166@end smallexample 3167 3168 3169@node Implementing THREADPRIVATE construct 3170@section Implementing THREADPRIVATE construct 3171 3172In _most_ cases we can map this directly to @code{__thread}. Except 3173that OMP allows constructors for C++ objects. We can either 3174refuse to support this (how often is it used?) or we can 3175implement something akin to .ctors. 3176 3177Even more ideally, this ctor feature is handled by extensions 3178to the main pthreads library. Failing that, we can have a set 3179of entry points to register ctor functions to be called. 3180 3181 3182 3183@node Implementing PRIVATE clause 3184@section Implementing PRIVATE clause 3185 3186In association with a PARALLEL, or within the lexical extent 3187of a PARALLEL block, the variable becomes a local variable in 3188the parallel subfunction. 3189 3190In association with FOR or SECTIONS blocks, create a new 3191automatic variable within the current function. This preserves 3192the semantic of new variable creation. 3193 3194 3195 3196@node Implementing FIRSTPRIVATE LASTPRIVATE COPYIN and COPYPRIVATE clauses 3197@section Implementing FIRSTPRIVATE LASTPRIVATE COPYIN and COPYPRIVATE clauses 3198 3199This seems simple enough for PARALLEL blocks. Create a private 3200struct for communicating between the parent and subfunction. 3201In the parent, copy in values for scalar and "small" structs; 3202copy in addresses for others TREE_ADDRESSABLE types. In the 3203subfunction, copy the value into the local variable. 3204 3205It is not clear what to do with bare FOR or SECTION blocks. 3206The only thing I can figure is that we do something like: 3207 3208@smallexample 3209#pragma omp for firstprivate(x) lastprivate(y) 3210for (int i = 0; i < n; ++i) 3211 body; 3212@end smallexample 3213 3214which becomes 3215 3216@smallexample 3217@{ 3218 int x = x, y; 3219 3220 // for stuff 3221 3222 if (i == n) 3223 y = y; 3224@} 3225@end smallexample 3226 3227where the "x=x" and "y=y" assignments actually have different 3228uids for the two variables, i.e. not something you could write 3229directly in C. Presumably this only makes sense if the "outer" 3230x and y are global variables. 3231 3232COPYPRIVATE would work the same way, except the structure 3233broadcast would have to happen via SINGLE machinery instead. 3234 3235 3236 3237@node Implementing REDUCTION clause 3238@section Implementing REDUCTION clause 3239 3240The private struct mentioned in the previous section should have 3241a pointer to an array of the type of the variable, indexed by the 3242thread's @var{team_id}. The thread stores its final value into the 3243array, and after the barrier, the master thread iterates over the 3244array to collect the values. 3245 3246 3247@node Implementing PARALLEL construct 3248@section Implementing PARALLEL construct 3249 3250@smallexample 3251 #pragma omp parallel 3252 @{ 3253 body; 3254 @} 3255@end smallexample 3256 3257becomes 3258 3259@smallexample 3260 void subfunction (void *data) 3261 @{ 3262 use data; 3263 body; 3264 @} 3265 3266 setup data; 3267 GOMP_parallel_start (subfunction, &data, num_threads); 3268 subfunction (&data); 3269 GOMP_parallel_end (); 3270@end smallexample 3271 3272@smallexample 3273 void GOMP_parallel_start (void (*fn)(void *), void *data, unsigned num_threads) 3274@end smallexample 3275 3276The @var{FN} argument is the subfunction to be run in parallel. 3277 3278The @var{DATA} argument is a pointer to a structure used to 3279communicate data in and out of the subfunction, as discussed 3280above with respect to FIRSTPRIVATE et al. 3281 3282The @var{NUM_THREADS} argument is 1 if an IF clause is present 3283and false, or the value of the NUM_THREADS clause, if 3284present, or 0. 3285 3286The function needs to create the appropriate number of 3287threads and/or launch them from the dock. It needs to 3288create the team structure and assign team ids. 3289 3290@smallexample 3291 void GOMP_parallel_end (void) 3292@end smallexample 3293 3294Tears down the team and returns us to the previous @code{omp_in_parallel()} state. 3295 3296 3297 3298@node Implementing FOR construct 3299@section Implementing FOR construct 3300 3301@smallexample 3302 #pragma omp parallel for 3303 for (i = lb; i <= ub; i++) 3304 body; 3305@end smallexample 3306 3307becomes 3308 3309@smallexample 3310 void subfunction (void *data) 3311 @{ 3312 long _s0, _e0; 3313 while (GOMP_loop_static_next (&_s0, &_e0)) 3314 @{ 3315 long _e1 = _e0, i; 3316 for (i = _s0; i < _e1; i++) 3317 body; 3318 @} 3319 GOMP_loop_end_nowait (); 3320 @} 3321 3322 GOMP_parallel_loop_static (subfunction, NULL, 0, lb, ub+1, 1, 0); 3323 subfunction (NULL); 3324 GOMP_parallel_end (); 3325@end smallexample 3326 3327@smallexample 3328 #pragma omp for schedule(runtime) 3329 for (i = 0; i < n; i++) 3330 body; 3331@end smallexample 3332 3333becomes 3334 3335@smallexample 3336 @{ 3337 long i, _s0, _e0; 3338 if (GOMP_loop_runtime_start (0, n, 1, &_s0, &_e0)) 3339 do @{ 3340 long _e1 = _e0; 3341 for (i = _s0, i < _e0; i++) 3342 body; 3343 @} while (GOMP_loop_runtime_next (&_s0, _&e0)); 3344 GOMP_loop_end (); 3345 @} 3346@end smallexample 3347 3348Note that while it looks like there is trickiness to propagating 3349a non-constant STEP, there isn't really. We're explicitly allowed 3350to evaluate it as many times as we want, and any variables involved 3351should automatically be handled as PRIVATE or SHARED like any other 3352variables. So the expression should remain evaluable in the 3353subfunction. We can also pull it into a local variable if we like, 3354but since its supposed to remain unchanged, we can also not if we like. 3355 3356If we have SCHEDULE(STATIC), and no ORDERED, then we ought to be 3357able to get away with no work-sharing context at all, since we can 3358simply perform the arithmetic directly in each thread to divide up 3359the iterations. Which would mean that we wouldn't need to call any 3360of these routines. 3361 3362There are separate routines for handling loops with an ORDERED 3363clause. Bookkeeping for that is non-trivial... 3364 3365 3366 3367@node Implementing ORDERED construct 3368@section Implementing ORDERED construct 3369 3370@smallexample 3371 void GOMP_ordered_start (void) 3372 void GOMP_ordered_end (void) 3373@end smallexample 3374 3375 3376 3377@node Implementing SECTIONS construct 3378@section Implementing SECTIONS construct 3379 3380A block as 3381 3382@smallexample 3383 #pragma omp sections 3384 @{ 3385 #pragma omp section 3386 stmt1; 3387 #pragma omp section 3388 stmt2; 3389 #pragma omp section 3390 stmt3; 3391 @} 3392@end smallexample 3393 3394becomes 3395 3396@smallexample 3397 for (i = GOMP_sections_start (3); i != 0; i = GOMP_sections_next ()) 3398 switch (i) 3399 @{ 3400 case 1: 3401 stmt1; 3402 break; 3403 case 2: 3404 stmt2; 3405 break; 3406 case 3: 3407 stmt3; 3408 break; 3409 @} 3410 GOMP_barrier (); 3411@end smallexample 3412 3413 3414@node Implementing SINGLE construct 3415@section Implementing SINGLE construct 3416 3417A block like 3418 3419@smallexample 3420 #pragma omp single 3421 @{ 3422 body; 3423 @} 3424@end smallexample 3425 3426becomes 3427 3428@smallexample 3429 if (GOMP_single_start ()) 3430 body; 3431 GOMP_barrier (); 3432@end smallexample 3433 3434while 3435 3436@smallexample 3437 #pragma omp single copyprivate(x) 3438 body; 3439@end smallexample 3440 3441becomes 3442 3443@smallexample 3444 datap = GOMP_single_copy_start (); 3445 if (datap == NULL) 3446 @{ 3447 body; 3448 data.x = x; 3449 GOMP_single_copy_end (&data); 3450 @} 3451 else 3452 x = datap->x; 3453 GOMP_barrier (); 3454@end smallexample 3455 3456 3457 3458@node Implementing OpenACC's PARALLEL construct 3459@section Implementing OpenACC's PARALLEL construct 3460 3461@smallexample 3462 void GOACC_parallel () 3463@end smallexample 3464 3465 3466 3467@c --------------------------------------------------------------------- 3468@c Reporting Bugs 3469@c --------------------------------------------------------------------- 3470 3471@node Reporting Bugs 3472@chapter Reporting Bugs 3473 3474Bugs in the GNU Offloading and Multi Processing Runtime Library should 3475be reported via @uref{http://gcc.gnu.org/bugzilla/, Bugzilla}. Please add 3476"openacc", or "openmp", or both to the keywords field in the bug 3477report, as appropriate. 3478 3479 3480 3481@c --------------------------------------------------------------------- 3482@c GNU General Public License 3483@c --------------------------------------------------------------------- 3484 3485@include gpl_v3.texi 3486 3487 3488 3489@c --------------------------------------------------------------------- 3490@c GNU Free Documentation License 3491@c --------------------------------------------------------------------- 3492 3493@include fdl.texi 3494 3495 3496 3497@c --------------------------------------------------------------------- 3498@c Funding Free Software 3499@c --------------------------------------------------------------------- 3500 3501@include funding.texi 3502 3503@c --------------------------------------------------------------------- 3504@c Index 3505@c --------------------------------------------------------------------- 3506 3507@node Library Index 3508@unnumbered Library Index 3509 3510@printindex cp 3511 3512@bye 3513