2.1 omp_get_active_level – Number of parallel regions

Description:

This function returns the nesting level for the active parallel blocks, which enclose the calling call.

C/C++

Prototype:

int omp_get_active_level(void);

Fortran:

Interface:

integer function omp_get_active_level()

See also:

omp_get_level – Obtain the current nesting level, omp_get_max_active_levels – Maximum number of active regions, omp_set_max_active_levels – Limits the number of active parallel regions

Reference:

OpenMP specification v4.5, Section 3.2.20.

2.2 omp_get_ancestor_thread_num – Ancestor thread ID

Description:

This function returns the thread identification number for the given nesting level of the current thread. For values of level outside zero to omp_get_level -1 is returned; if level is omp_get_level the result is identical to omp_get_thread_num.

C/C++

Prototype:

int omp_get_ancestor_thread_num(int level);

Fortran:

Interface:	integer function omp_get_ancestor_thread_num(level)
	integer level

See also:

omp_get_level – Obtain the current nesting level, omp_get_thread_num – Current thread ID, omp_get_team_size – Number of threads in a team

Reference:

OpenMP specification v4.5, Section 3.2.18.

2.3 omp_get_cancellation – Whether cancellation support is enabled

Description:

This function returns true if cancellation is activated, false otherwise. Here, true and false represent their language-specific counterparts. Unless OMP_CANCELLATION is set true, cancellations are deactivated.

C/C++:

Prototype:

int omp_get_cancellation(void);

Fortran:

Interface:

logical function omp_get_cancellation()

See also:

OMP_CANCELLATION – Set whether cancellation is activated

Reference:

OpenMP specification v4.5, Section 3.2.9.

2.4 omp_get_default_device – Get the default device for target regions

Description:

Get the default device for target regions without device clause.

C/C++:

Prototype:

int omp_get_default_device(void);

Fortran:

Interface:

integer function omp_get_default_device()

See also:

OMP_DEFAULT_DEVICE – Set the device used in target regions, omp_set_default_device – Set the default device for target regions

Reference:

OpenMP specification v4.5, Section 3.2.30.

2.5 omp_get_dynamic – Dynamic teams setting

Description:

This function returns true if enabled, false otherwise. Here, true and false represent their language-specific counterparts.

The dynamic team setting may be initialized at startup by the OMP_DYNAMIC environment variable or at runtime using omp_set_dynamic. If undefined, dynamic adjustment is disabled by default.

C/C++:

Prototype:

int omp_get_dynamic(void);

Fortran:

Interface:

logical function omp_get_dynamic()

See also:

omp_set_dynamic – Enable/disable dynamic teams, OMP_DYNAMIC – Dynamic adjustment of threads

Reference:

OpenMP specification v4.5, Section 3.2.8.

2.6 omp_get_level – Obtain the current nesting level

Description:

This function returns the nesting level for the parallel blocks, which enclose the calling call.

C/C++

Prototype:

int omp_get_level(void);

Fortran:

Interface:

integer function omp_level()

See also:

omp_get_active_level – Number of parallel regions

Reference:

OpenMP specification v4.5, Section 3.2.17.

2.7 omp_get_max_active_levels – Maximum number of active regions

Description:

This function obtains the maximum allowed number of nested, active parallel regions.

C/C++

Prototype:

int omp_get_max_active_levels(void);

Fortran:

Interface:

integer function omp_get_max_active_levels()

See also:

omp_set_max_active_levels – Limits the number of active parallel regions, omp_get_active_level – Number of parallel regions

Reference:

OpenMP specification v4.5, Section 3.2.16.

2.8 omp_get_max_task_priority – Maximum priority value

that can be set for tasks.

Description:

This function obtains the maximum allowed priority number for tasks.

C/C++

Prototype:

int omp_get_max_task_priority(void);

Fortran:

Interface:

integer function omp_get_max_task_priority()

Reference:

OpenMP specification v4.5, Section 3.2.29.

2.9 omp_get_max_threads – Maximum number of threads of parallel region

Description:

Return the maximum number of threads used for the current parallel region that does not use the clause num_threads.

C/C++:

Prototype:

int omp_get_max_threads(void);

Fortran:

Interface:

integer function omp_get_max_threads()

See also:

omp_set_num_threads – Set upper team size limit, omp_set_dynamic – Enable/disable dynamic teams, omp_get_thread_limit – Maximum number of threads

Reference:

OpenMP specification v4.5, Section 3.2.3.

2.10 omp_get_nested – Nested parallel regions

Description:

This function returns true if nested parallel regions are enabled, false otherwise. Here, true and false represent their language-specific counterparts.

Nested parallel regions may be initialized at startup by the OMP_NESTED environment variable or at runtime using omp_set_nested. If undefined, nested parallel regions are disabled by default.

C/C++:

Prototype:

int omp_get_nested(void);

Fortran:

Interface:

logical function omp_get_nested()

See also:

omp_set_nested – Enable/disable nested parallel regions, OMP_NESTED – Nested parallel regions

Reference:

OpenMP specification v4.5, Section 3.2.11.

2.11 omp_get_num_devices – Number of target devices

Description:

Returns the number of target devices.

C/C++:

Prototype:

int omp_get_num_devices(void);

Fortran:

Interface:

integer function omp_get_num_devices()

Reference:

OpenMP specification v4.5, Section 3.2.31.

2.12 omp_get_num_procs – Number of processors online

Description:

Returns the number of processors online on that device.

C/C++:

Prototype:

int omp_get_num_procs(void);

Fortran:

Interface:

integer function omp_get_num_procs()

Reference:

OpenMP specification v4.5, Section 3.2.5.

2.13 omp_get_num_teams – Number of teams

Description:

Returns the number of teams in the current team region.

C/C++:

Prototype:

int omp_get_num_teams(void);

Fortran:

Interface:

integer function omp_get_num_teams()

Reference:

OpenMP specification v4.5, Section 3.2.32.

2.14 omp_get_num_threads – Size of the active team

Description:

Returns the number of threads in the current team. In a sequential section of the program omp_get_num_threads returns 1.

The default team size may be initialized at startup by the OMP_NUM_THREADS environment variable. At runtime, the size of the current team may be set either by the NUM_THREADS clause or by omp_set_num_threads. If none of the above were used to define a specific value and OMP_DYNAMIC is disabled, one thread per CPU online is used.

C/C++:

Prototype:

int omp_get_num_threads(void);

Fortran:

Interface:

integer function omp_get_num_threads()

See also:

omp_get_max_threads – Maximum number of threads of parallel region, omp_set_num_threads – Set upper team size limit, OMP_NUM_THREADS – Specifies the number of threads to use

Reference:

OpenMP specification v4.5, Section 3.2.2.

2.15 omp_get_proc_bind – Whether threads may be moved between CPUs

Description:

This functions returns the currently active thread affinity policy, which is set via OMP_PROC_BIND. Possible values are omp_proc_bind_false, omp_proc_bind_true, omp_proc_bind_master, omp_proc_bind_close and omp_proc_bind_spread.

C/C++:

Prototype:

omp_proc_bind_t omp_get_proc_bind(void);

Fortran:

Interface:

integer(kind=omp_proc_bind_kind) function omp_get_proc_bind()

See also:

OMP_PROC_BIND – Whether threads may be moved between CPUs, OMP_PLACES – Specifies on which CPUs the threads should be placed, GOMP_CPU_AFFINITY – Bind threads to specific CPUs,

Reference:

OpenMP specification v4.5, Section 3.2.22.

2.16 omp_get_schedule – Obtain the runtime scheduling method

Description:

Obtain the runtime scheduling method. The kind argument will be set to the value omp_sched_static, omp_sched_dynamic, omp_sched_guided or omp_sched_auto. The second argument, chunk_size, is set to the chunk size.

C/C++

Prototype:

void omp_get_schedule(omp_sched_t *kind, int *chunk_size);

Fortran:

Interface:	subroutine omp_get_schedule(kind, chunk_size)
	integer(kind=omp_sched_kind) kind
	integer chunk_size

See also:

omp_set_schedule – Set the runtime scheduling method, OMP_SCHEDULE – How threads are scheduled

Reference:

OpenMP specification v4.5, Section 3.2.13.

2.17 omp_get_team_num – Get team number

Description:

Returns the team number of the calling thread.

C/C++:

Prototype:

int omp_get_team_num(void);

Fortran:

Interface:

integer function omp_get_team_num()

Reference:

OpenMP specification v4.5, Section 3.2.33.

2.18 omp_get_team_size – Number of threads in a team

Description:

This function returns the number of threads in a thread team to which either the current thread or its ancestor belongs. For values of level outside zero to omp_get_level, -1 is returned; if level is zero, 1 is returned, and for omp_get_level, the result is identical to omp_get_num_threads.

C/C++:

Prototype:

int omp_get_team_size(int level);

Fortran:

Interface:	integer function omp_get_team_size(level)
	integer level

See also:

omp_get_num_threads – Size of the active team, omp_get_level – Obtain the current nesting level, omp_get_ancestor_thread_num – Ancestor thread ID

Reference:

OpenMP specification v4.5, Section 3.2.19.

2.19 omp_get_thread_limit – Maximum number of threads

Description:

Return the maximum number of threads of the program.

C/C++:

Prototype:

int omp_get_thread_limit(void);

Fortran:

Interface:

integer function omp_get_thread_limit()

See also:

omp_get_max_threads – Maximum number of threads of parallel region, OMP_THREAD_LIMIT – Set the maximum number of threads

Reference:

OpenMP specification v4.5, Section 3.2.14.

2.20 omp_get_thread_num – Current thread ID

Description:

Returns a unique thread identification number within the current team. In a sequential parts of the program, omp_get_thread_num always returns 0. In parallel regions the return value varies from 0 to omp_get_num_threads-1 inclusive. The return value of the master thread of a team is always 0.

C/C++:

Prototype:

int omp_get_thread_num(void);

Fortran:

Interface:

integer function omp_get_thread_num()

See also:

omp_get_num_threads – Size of the active team, omp_get_ancestor_thread_num – Ancestor thread ID

Reference:

OpenMP specification v4.5, Section 3.2.4.

2.21 omp_in_parallel – Whether a parallel region is active

Description:

This function returns true if currently running in parallel, false otherwise. Here, true and false represent their language-specific counterparts.

C/C++:

Prototype:

int omp_in_parallel(void);

Fortran:

Interface:

logical function omp_in_parallel()

Reference:

OpenMP specification v4.5, Section 3.2.6.

2.22 omp_in_final – Whether in final or included task region

Description:

This function returns true if currently running in a final or included task region, false otherwise. Here, true and false represent their language-specific counterparts.

C/C++:

Prototype:

int omp_in_final(void);

Fortran:

Interface:

logical function omp_in_final()

Reference:

OpenMP specification v4.5, Section 3.2.21.

2.23 omp_is_initial_device – Whether executing on the host device

Description:

This function returns true if currently running on the host device, false otherwise. Here, true and false represent their language-specific counterparts.

C/C++:

Prototype:

int omp_is_initial_device(void);

Fortran:

Interface:

logical function omp_is_initial_device()

Reference:

OpenMP specification v4.5, Section 3.2.34.

2.24 omp_set_default_device – Set the default device for target regions

Description:

Set the default device for target regions without device clause. The argument shall be a nonnegative device number.

C/C++:

Prototype:

void omp_set_default_device(int device_num);

Fortran:

Interface:	subroutine omp_set_default_device(device_num)
	integer device_num

See also:

OMP_DEFAULT_DEVICE – Set the device used in target regions, omp_get_default_device – Get the default device for target regions

Reference:

OpenMP specification v4.5, Section 3.2.29.

2.25 omp_set_dynamic – Enable/disable dynamic teams

Description:

Enable or disable the dynamic adjustment of the number of threads within a team. The function takes the language-specific equivalent of true and false, where true enables dynamic adjustment of team sizes and false disables it.

C/C++:

Prototype:

void omp_set_dynamic(int dynamic_threads);

Fortran:

Interface:	subroutine omp_set_dynamic(dynamic_threads)
	logical, intent(in) :: dynamic_threads

See also:

OMP_DYNAMIC – Dynamic adjustment of threads, omp_get_dynamic – Dynamic teams setting

Reference:

OpenMP specification v4.5, Section 3.2.7.

2.26 omp_set_max_active_levels – Limits the number of active parallel regions

Description:

This function limits the maximum allowed number of nested, active parallel regions.

C/C++

Prototype:

void omp_set_max_active_levels(int max_levels);

Fortran:

Interface:	subroutine omp_set_max_active_levels(max_levels)
	integer max_levels

See also:

omp_get_max_active_levels – Maximum number of active regions, omp_get_active_level – Number of parallel regions

Reference:

OpenMP specification v4.5, Section 3.2.15.

2.27 omp_set_nested – Enable/disable nested parallel regions

Description:

Enable or disable nested parallel regions, i.e., whether team members are allowed to create new teams. The function takes the language-specific equivalent of true and false, where true enables dynamic adjustment of team sizes and false disables it.

C/C++:

Prototype:

void omp_set_nested(int nested);

Fortran:

Interface:	subroutine omp_set_nested(nested)
	logical, intent(in) :: nested

See also:

OMP_NESTED – Nested parallel regions, omp_get_nested – Nested parallel regions

Reference:

OpenMP specification v4.5, Section 3.2.10.

2.28 omp_set_num_threads – Set upper team size limit

Description:

Specifies the number of threads used by default in subsequent parallel sections, if those do not specify a num_threads clause. The argument of omp_set_num_threads shall be a positive integer.

C/C++:

Prototype:

void omp_set_num_threads(int num_threads);

Fortran:

Interface:	subroutine omp_set_num_threads(num_threads)
	integer, intent(in) :: num_threads

See also:

OMP_NUM_THREADS – Specifies the number of threads to use, omp_get_num_threads – Size of the active team, omp_get_max_threads – Maximum number of threads of parallel region

Reference:

OpenMP specification v4.5, Section 3.2.1.

2.29 omp_set_schedule – Set the runtime scheduling method

Description:

Sets the runtime scheduling method. The kind argument can have the value omp_sched_static, omp_sched_dynamic, omp_sched_guided or omp_sched_auto. Except for omp_sched_auto, the chunk size is set to the value of chunk_size if positive, or to the default value if zero or negative. For omp_sched_auto the chunk_size argument is ignored.

C/C++

Prototype:

void omp_set_schedule(omp_sched_t kind, int chunk_size);

Fortran:

Interface:	subroutine omp_set_schedule(kind, chunk_size)
	integer(kind=omp_sched_kind) kind
	integer chunk_size

See also:

omp_get_schedule – Obtain the runtime scheduling method OMP_SCHEDULE – How threads are scheduled

Reference:

OpenMP specification v4.5, Section 3.2.12.

2.30 omp_init_lock – Initialize simple lock

Description:

Initialize a simple lock. After initialization, the lock is in an unlocked state.

C/C++:

Prototype:

void omp_init_lock(omp_lock_t *lock);

Fortran:

Interface:	subroutine omp_init_lock(svar)
	integer(omp_lock_kind), intent(out) :: svar

See also:

omp_destroy_lock – Destroy simple lock

Reference:

OpenMP specification v4.5, Section 3.3.1.

2.31 omp_set_lock – Wait for and set simple lock

Description:

Before setting a simple lock, the lock variable must be initialized by omp_init_lock. The calling thread is blocked until the lock is available. If the lock is already held by the current thread, a deadlock occurs.

C/C++:

Prototype:

void omp_set_lock(omp_lock_t *lock);

Fortran:

Interface:	subroutine omp_set_lock(svar)
	integer(omp_lock_kind), intent(inout) :: svar

See also:

omp_init_lock – Initialize simple lock, omp_test_lock – Test and set simple lock if available, omp_unset_lock – Unset simple lock

Reference:

OpenMP specification v4.5, Section 3.3.4.

2.32 omp_test_lock – Test and set simple lock if available

Description:

Before setting a simple lock, the lock variable must be initialized by omp_init_lock. Contrary to omp_set_lock, omp_test_lock does not block if the lock is not available. This function returns true upon success, false otherwise. Here, true and false represent their language-specific counterparts.

C/C++:

Prototype:

int omp_test_lock(omp_lock_t *lock);

Fortran:

Interface:	logical function omp_test_lock(svar)
	integer(omp_lock_kind), intent(inout) :: svar

See also:

omp_init_lock – Initialize simple lock, omp_set_lock – Wait for and set simple lock, omp_set_lock – Wait for and set simple lock

Reference:

OpenMP specification v4.5, Section 3.3.6.

2.33 omp_unset_lock – Unset simple lock

Description:

A simple lock about to be unset must have been locked by omp_set_lock or omp_test_lock before. In addition, the lock must be held by the thread calling omp_unset_lock. Then, the lock becomes unlocked. If one or more threads attempted to set the lock before, one of them is chosen to, again, set the lock to itself.

C/C++:

Prototype:

void omp_unset_lock(omp_lock_t *lock);

Fortran:

Interface:	subroutine omp_unset_lock(svar)
	integer(omp_lock_kind), intent(inout) :: svar

See also:

omp_set_lock – Wait for and set simple lock, omp_test_lock – Test and set simple lock if available

Reference:

OpenMP specification v4.5, Section 3.3.5.

2.34 omp_destroy_lock – Destroy simple lock

Description:

Destroy a simple lock. In order to be destroyed, a simple lock must be in the unlocked state.

C/C++:

Prototype:

void omp_destroy_lock(omp_lock_t *lock);

Fortran:

Interface:	subroutine omp_destroy_lock(svar)
	integer(omp_lock_kind), intent(inout) :: svar

See also:

omp_init_lock – Initialize simple lock

Reference:

OpenMP specification v4.5, Section 3.3.3.

2.35 omp_init_nest_lock – Initialize nested lock

Description:

Initialize a nested lock. After initialization, the lock is in an unlocked state and the nesting count is set to zero.

C/C++:

Prototype:

void omp_init_nest_lock(omp_nest_lock_t *lock);

Fortran:

Interface:	subroutine omp_init_nest_lock(nvar)
	integer(omp_nest_lock_kind), intent(out) :: nvar

See also:

omp_destroy_nest_lock – Destroy nested lock

Reference:

OpenMP specification v4.5, Section 3.3.1.

2.36 omp_set_nest_lock – Wait for and set nested lock

Description:

Before setting a nested lock, the lock variable must be initialized by omp_init_nest_lock. The calling thread is blocked until the lock is available. If the lock is already held by the current thread, the nesting count for the lock is incremented.

C/C++:

Prototype:

void omp_set_nest_lock(omp_nest_lock_t *lock);

Fortran:

Interface:	subroutine omp_set_nest_lock(nvar)
	integer(omp_nest_lock_kind), intent(inout) :: nvar

See also:

omp_init_nest_lock – Initialize nested lock, omp_unset_nest_lock – Unset nested lock

Reference:

OpenMP specification v4.5, Section 3.3.4.

2.37 omp_test_nest_lock – Test and set nested lock if available

Description:

Before setting a nested lock, the lock variable must be initialized by omp_init_nest_lock. Contrary to omp_set_nest_lock, omp_test_nest_lock does not block if the lock is not available. If the lock is already held by the current thread, the new nesting count is returned. Otherwise, the return value equals zero.

C/C++:

Prototype:

int omp_test_nest_lock(omp_nest_lock_t *lock);

Fortran:

Interface:	logical function omp_test_nest_lock(nvar)
	integer(omp_nest_lock_kind), intent(inout) :: nvar

See also:

omp_init_lock – Initialize simple lock, omp_set_lock – Wait for and set simple lock, omp_set_lock – Wait for and set simple lock

Reference:

OpenMP specification v4.5, Section 3.3.6.

2.38 omp_unset_nest_lock – Unset nested lock

Description:

A nested lock about to be unset must have been locked by omp_set_nested_lock or omp_test_nested_lock before. In addition, the lock must be held by the thread calling omp_unset_nested_lock. If the nesting count drops to zero, the lock becomes unlocked. If one ore more threads attempted to set the lock before, one of them is chosen to, again, set the lock to itself.

C/C++:

Prototype:

void omp_unset_nest_lock(omp_nest_lock_t *lock);

Fortran:

Interface:	subroutine omp_unset_nest_lock(nvar)
	integer(omp_nest_lock_kind), intent(inout) :: nvar

See also:

omp_set_nest_lock – Wait for and set nested lock

Reference:

OpenMP specification v4.5, Section 3.3.5.

2.39 omp_destroy_nest_lock – Destroy nested lock

Description:

Destroy a nested lock. In order to be destroyed, a nested lock must be in the unlocked state and its nesting count must equal zero.

C/C++:

Prototype:

void omp_destroy_nest_lock(omp_nest_lock_t *);

Fortran:

Interface:	subroutine omp_destroy_nest_lock(nvar)
	integer(omp_nest_lock_kind), intent(inout) :: nvar

See also:

omp_init_lock – Initialize simple lock

Reference:

OpenMP specification v4.5, Section 3.3.3.

2.40 omp_get_wtick – Get timer precision

Description:

Gets the timer precision, i.e., the number of seconds between two successive clock ticks.

C/C++:

Prototype:

double omp_get_wtick(void);

Fortran:

Interface:

double precision function omp_get_wtick()

See also:

omp_get_wtime – Elapsed wall clock time

Reference:

OpenMP specification v4.5, Section 3.4.2.

2.41 omp_get_wtime – Elapsed wall clock time

Description:

Elapsed wall clock time in seconds. The time is measured per thread, no guarantee can be made that two distinct threads measure the same time. Time is measured from some "time in the past", which is an arbitrary time guaranteed not to change during the execution of the program.

C/C++:

Prototype:

double omp_get_wtime(void);

Fortran:

Interface:

double precision function omp_get_wtime()

See also:

omp_get_wtick – Get timer precision

Reference:

OpenMP specification v4.5, Section 3.4.1.

3.1 OMP_CANCELLATION – Set whether cancellation is activated

Description:

If set to TRUE, the cancellation is activated. If set to FALSE or if unset, cancellation is disabled and the cancel construct is ignored.

See also:

omp_get_cancellation – Whether cancellation support is enabled

Reference:

OpenMP specification v4.5, Section 4.11

3.2 OMP_DISPLAY_ENV – Show OpenMP version and environment variables

Description:

If set to TRUE, the OpenMP version number and the values associated with the OpenMP environment variables are printed to stderr. If set to VERBOSE, it additionally shows the value of the environment variables which are GNU extensions. If undefined or set to FALSE, this information will not be shown.

Reference:

OpenMP specification v4.5, Section 4.12

3.3 OMP_DEFAULT_DEVICE – Set the device used in target regions

Description:

Set to choose the device which is used in a target region, unless the value is overridden by omp_set_default_device or by a device clause. The value shall be the nonnegative device number. If no device with the given device number exists, the code is executed on the host. If unset, device number 0 will be used.

See also:

omp_get_default_device – Get the default device for target regions, omp_set_default_device – Set the default device for target regions,

Reference:

OpenMP specification v4.5, Section 4.13

3.4 OMP_DYNAMIC – Dynamic adjustment of threads

Description:

Enable or disable the dynamic adjustment of the number of threads within a team. The value of this environment variable shall be TRUE or FALSE. If undefined, dynamic adjustment is disabled by default.

See also:

omp_set_dynamic – Enable/disable dynamic teams

Reference:

OpenMP specification v4.5, Section 4.3

3.5 OMP_MAX_ACTIVE_LEVELS – Set the maximum number of nested parallel regions

Description:

Specifies the initial value for the maximum number of nested parallel regions. The value of this variable shall be a positive integer. If undefined, the number of active levels is unlimited.

See also:

omp_set_max_active_levels – Limits the number of active parallel regions

Reference:

OpenMP specification v4.5, Section 4.9

3.6 OMP_MAX_TASK_PRIORITY – Set the maximum priority

number that can be set for a task.

Description:

Specifies the initial value for the maximum priority value that can be set for a task. The value of this variable shall be a non-negative integer, and zero is allowed. If undefined, the default priority is 0.

See also:

omp_get_max_task_priority – Maximum priority value

Reference:

OpenMP specification v4.5, Section 4.14

3.7 OMP_NESTED – Nested parallel regions

Description:

Enable or disable nested parallel regions, i.e., whether team members are allowed to create new teams. The value of this environment variable shall be TRUE or FALSE. If undefined, nested parallel regions are disabled by default.

See also:

omp_set_nested – Enable/disable nested parallel regions

Reference:

OpenMP specification v4.5, Section 4.6

3.8 OMP_NUM_THREADS – Specifies the number of threads to use

Description:

Specifies the default number of threads to use in parallel regions. The value of this variable shall be a comma-separated list of positive integers; the value specified the number of threads to use for the corresponding nested level. If undefined one thread per CPU is used.

See also:

omp_set_num_threads – Set upper team size limit

Reference:

OpenMP specification v4.5, Section 4.2

3.9 OMP_PROC_BIND – Whether threads may be moved between CPUs

Description:

Specifies whether threads may be moved between processors. If set to TRUE, OpenMP threads should not be moved; if set to FALSE they may be moved. Alternatively, a comma separated list with the values MASTER, CLOSE and SPREAD can be used to specify the thread affinity policy for the corresponding nesting level. With MASTER the worker threads are in the same place partition as the master thread. With CLOSE those are kept close to the master thread in contiguous place partitions. And with SPREAD a sparse distribution across the place partitions is used.

When undefined, OMP_PROC_BIND defaults to TRUE when OMP_PLACES or GOMP_CPU_AFFINITY is set and FALSE otherwise.

See also:

OMP_PLACES – Specifies on which CPUs the threads should be placed, GOMP_CPU_AFFINITY – Bind threads to specific CPUs, omp_get_proc_bind – Whether threads may be moved between CPUs

Reference:

OpenMP specification v4.5, Section 4.4

3.10 OMP_PLACES – Specifies on which CPUs the threads should be placed

Description:

The thread placement can be either specified using an abstract name or by an explicit list of the places. The abstract names threads, cores and sockets can be optionally followed by a positive number in parentheses, which denotes the how many places shall be created. With threads each place corresponds to a single hardware thread; cores to a single core with the corresponding number of hardware threads; and with sockets the place corresponds to a single socket. The resulting placement can be shown by setting the OMP_DISPLAY_ENV environment variable.

Alternatively, the placement can be specified explicitly as comma-separated list of places. A place is specified by set of nonnegative numbers in curly braces, denoting the denoting the hardware threads. The hardware threads belonging to a place can either be specified as comma-separated list of nonnegative thread numbers or using an interval. Multiple places can also be either specified by a comma-separated list of places or by an interval. To specify an interval, a colon followed by the count is placed after after the hardware thread number or the place. Optionally, the length can be followed by a colon and the stride number – otherwise a unit stride is assumed. For instance, the following specifies the same places list: "{0,1,2}, {3,4,6}, {7,8,9}, {10,11,12}"; "{0:3}, {3:3}, {7:3}, {10:3}"; and "{0:2}:4:3".

If OMP_PLACES and GOMP_CPU_AFFINITY are unset and OMP_PROC_BIND is either unset or false, threads may be moved between CPUs following no placement policy.

See also:

OMP_PROC_BIND – Whether threads may be moved between CPUs, GOMP_CPU_AFFINITY – Bind threads to specific CPUs, omp_get_proc_bind – Whether threads may be moved between CPUs, OMP_DISPLAY_ENV – Show OpenMP version and environment variables

Reference:

OpenMP specification v4.5, Section 4.5

3.11 OMP_STACKSIZE – Set default thread stack size

Description:

Set the default thread stack size in kilobytes, unless the number is suffixed by B, K, M or G, in which case the size is, respectively, in bytes, kilobytes, megabytes or gigabytes. This is different from pthread_attr_setstacksize which gets the number of bytes as an argument. If the stack size cannot be set due to system constraints, an error is reported and the initial stack size is left unchanged. If undefined, the stack size is system dependent.

Reference:

OpenMP specification v4.5, Section 4.7

3.12 OMP_SCHEDULE – How threads are scheduled

Description:

Allows to specify schedule type and chunk size. The value of the variable shall have the form: type[,chunk] where type is one of static, dynamic, guided or auto The optional chunk size shall be a positive integer. If undefined, dynamic scheduling and a chunk size of 1 is used.

See also:

omp_set_schedule – Set the runtime scheduling method

Reference:

OpenMP specification v4.5, Sections 2.7.1.1 and 4.1

3.13 OMP_THREAD_LIMIT – Set the maximum number of threads

Description:

Specifies the number of threads to use for the whole program. The value of this variable shall be a positive integer. If undefined, the number of threads is not limited.

See also:

OMP_NUM_THREADS – Specifies the number of threads to use, omp_get_thread_limit – Maximum number of threads

Reference:

OpenMP specification v4.5, Section 4.10

3.14 OMP_WAIT_POLICY – How waiting threads are handled

Description:

Specifies whether waiting threads should be active or passive. If the value is PASSIVE, waiting threads should not consume CPU power while waiting; while the value is ACTIVE specifies that they should. If undefined, threads wait actively for a short time before waiting passively.

See also:

GOMP_SPINCOUNT – Set the busy-wait spin count

Reference:

OpenMP specification v4.5, Section 4.8

3.15 GOMP_CPU_AFFINITY – Bind threads to specific CPUs

Description:

Binds threads to specific CPUs. The variable should contain a space-separated or comma-separated list of CPUs. This list may contain different kinds of entries: either single CPU numbers in any order, a range of CPUs (M-N) or a range with some stride (M-N:S). CPU numbers are zero based. For example, GOMP_CPU_AFFINITY="0 3 1-2 4-15:2" will bind the initial thread to CPU 0, the second to CPU 3, the third to CPU 1, the fourth to CPU 2, the fifth to CPU 4, the sixth through tenth to CPUs 6, 8, 10, 12, and 14 respectively and then start assigning back from the beginning of the list. GOMP_CPU_AFFINITY=0 binds all threads to CPU 0.

There is no libgomp library routine to determine whether a CPU affinity specification is in effect. As a workaround, language-specific library functions, e.g., getenv in C or GET_ENVIRONMENT_VARIABLE in Fortran, may be used to query the setting of the GOMP_CPU_AFFINITY environment variable. A defined CPU affinity on startup cannot be changed or disabled during the runtime of the application.

If both GOMP_CPU_AFFINITY and OMP_PROC_BIND are set, OMP_PROC_BIND has a higher precedence. If neither has been set and OMP_PROC_BIND is unset, or when OMP_PROC_BIND is set to FALSE, the host system will handle the assignment of threads to CPUs.

See also:

OMP_PLACES – Specifies on which CPUs the threads should be placed, OMP_PROC_BIND – Whether threads may be moved between CPUs

3.16 GOMP_DEBUG – Enable debugging output

Description:

Enable debugging output. The variable should be set to 0 (disabled, also the default if not set), or 1 (enabled).

If enabled, some debugging output will be printed during execution. This is currently not specified in more detail, and subject to change.

3.17 GOMP_STACKSIZE – Set default thread stack size

Description:

Set the default thread stack size in kilobytes. This is different from pthread_attr_setstacksize which gets the number of bytes as an argument. If the stack size cannot be set due to system constraints, an error is reported and the initial stack size is left unchanged. If undefined, the stack size is system dependent.

See also:

OMP_STACKSIZE – Set default thread stack size

Reference:

GCC Patches Mailinglist, GCC Patches Mailinglist

3.18 GOMP_SPINCOUNT – Set the busy-wait spin count

Description:

Determines how long a threads waits actively with consuming CPU power before waiting passively without consuming CPU power. The value may be either INFINITE, INFINITY to always wait actively or an integer which gives the number of spins of the busy-wait loop. The integer may optionally be followed by the following suffixes acting as multiplication factors: k (kilo, thousand), M (mega, million), G (giga, billion), or T (tera, trillion). If undefined, 0 is used when OMP_WAIT_POLICY is PASSIVE, 300,000 is used when OMP_WAIT_POLICY is undefined and 30 billion is used when OMP_WAIT_POLICY is ACTIVE. If there are more OpenMP threads than available CPUs, 1000 and 100 spins are used for OMP_WAIT_POLICY being ACTIVE or undefined, respectively; unless the GOMP_SPINCOUNT is lower or OMP_WAIT_POLICY is PASSIVE.

See also:

OMP_WAIT_POLICY – How waiting threads are handled

3.19 GOMP_RTEMS_THREAD_POOLS – Set the RTEMS specific thread pools

Description:

This environment variable is only used on the RTEMS real-time operating system. It determines the scheduler instance specific thread pools. The format for GOMP_RTEMS_THREAD_POOLS is a list of optional <thread-pool-count>[$<priority>]@<scheduler-name> configurations separated by : where:

• <thread-pool-count> is the thread pool count for this scheduler instance.
• $<priority> is an optional priority for the worker threads of a thread pool according to pthread_setschedparam. In case a priority value is omitted, then a worker thread will inherit the priority of the OpenMP master thread that created it. The priority of the worker thread is not changed after creation, even if a new OpenMP master thread using the worker has a different priority.
• @<scheduler-name> is the scheduler instance name according to the RTEMS application configuration.

In case no thread pool configuration is specified for a scheduler instance, then each OpenMP master thread of this scheduler instance will use its own dynamically allocated thread pool. To limit the worker thread count of the thread pools, each OpenMP master thread must call omp_set_num_threads.

Example:

Lets suppose we have three scheduler instances IO, WRK0, and WRK1 with GOMP_RTEMS_THREAD_POOLS set to "1@WRK0:3$4@WRK1". Then there are no thread pool restrictions for scheduler instance IO. In the scheduler instance WRK0 there is one thread pool available. Since no priority is specified for this scheduler instance, the worker thread inherits the priority of the OpenMP master thread that created it. In the scheduler instance WRK1 there are three thread pools available and their worker threads run at priority four.

5.1 acc_get_num_devices – Get number of devices for given device type

Description

This function returns a value indicating the number of devices available for the device type specified in devicetype.

C/C++:

Prototype:

int acc_get_num_devices(acc_device_t devicetype);

Fortran:

Interface:	integer function acc_get_num_devices(devicetype)
	integer(kind=acc_device_kind) devicetype

Reference:

OpenACC specification v2.6, section 3.2.1.

5.2 acc_set_device_type – Set type of device accelerator to use.

Description

This function indicates to the runtime library which device type, specified in devicetype, to use when executing a parallel or kernels region.

C/C++:

Prototype:

acc_set_device_type(acc_device_t devicetype);

Fortran:

Interface:	subroutine acc_set_device_type(devicetype)
	integer(kind=acc_device_kind) devicetype

Reference:

OpenACC specification v2.6, section 3.2.2.

5.3 acc_get_device_type – Get type of device accelerator to be used.

Description

This function returns what device type will be used when executing a parallel or kernels region.

This function returns acc_device_none if acc_get_device_type is called from acc_ev_device_init_start, acc_ev_device_init_end callbacks of the OpenACC Profiling Interface (OpenACC Profiling Interface), that is, if the device is currently being initialized.

C/C++:

Prototype:

acc_device_t acc_get_device_type(void);

Fortran:

Interface:	function acc_get_device_type(void)
	integer(kind=acc_device_kind) acc_get_device_type

Reference:

OpenACC specification v2.6, section 3.2.3.

5.4 acc_set_device_num – Set device number to use.

Description

This function will indicate to the runtime which device number, specified by devicenum, associated with the specified device type devicetype.

C/C++:

Prototype:

acc_set_device_num(int devicenum, acc_device_t devicetype);

Fortran:

Interface:	subroutine acc_set_device_num(devicenum, devicetype)
	integer devicenum
	integer(kind=acc_device_kind) devicetype

Reference:

OpenACC specification v2.6, section 3.2.4.

5.5 acc_get_device_num – Get device number to be used.

Description

This function returns which device number associated with the specified device type devicetype, will be used when executing a parallel or kernels region.

C/C++:

Prototype:

int acc_get_device_num(acc_device_t devicetype);

Fortran:

Interface:	function acc_get_device_num(devicetype)
	integer(kind=acc_device_kind) devicetype
	integer acc_get_device_num

Reference:

OpenACC specification v2.6, section 3.2.5.

5.6 acc_get_property – Get device property.

Description

These routines return the value of the specified property for the device being queried according to devicenum and devicetype. Integer-valued and string-valued properties are returned by acc_get_property and acc_get_property_string respectively. The Fortran acc_get_property_string subroutine returns the string retrieved in its fourth argument while the remaining entry points are functions, which pass the return value as their result.

Note for Fortran, only: the OpenACC technical committee corrected and, hence, modified the interface introduced in OpenACC 2.6. The kind-value parameter acc_device_property has been renamed to acc_device_property_kind for consistency and the return type of the acc_get_property function is now a c_size_t integer instead of a acc_device_property integer. The parameter acc_device_property will continue to be provided, but might be removed in a future version of GCC.

C/C++:

Prototype:	size_t acc_get_property(int devicenum, acc_device_t devicetype, acc_device_property_t property);
Prototype:	const char *acc_get_property_string(int devicenum, acc_device_t devicetype, acc_device_property_t property);

Fortran:

Interface:	function acc_get_property(devicenum, devicetype, property)
Interface:	subroutine acc_get_property_string(devicenum, devicetype, property, string)
	use ISO_C_Binding, only: c_size_t
	integer devicenum
	integer(kind=acc_device_kind) devicetype
	integer(kind=acc_device_property_kind) property
	integer(kind=c_size_t) acc_get_property
	character(*) string

Reference:

OpenACC specification v2.6, section 3.2.6.

5.7 acc_async_test – Test for completion of a specific asynchronous operation.

Description

This function tests for completion of the asynchronous operation specified in arg. In C/C++, a non-zero value will be returned to indicate the specified asynchronous operation has completed. While Fortran will return a true. If the asynchronous operation has not completed, C/C++ returns a zero and Fortran returns a false.

C/C++:

Prototype:

int acc_async_test(int arg);

Fortran:

Interface:	function acc_async_test(arg)
	integer(kind=acc_handle_kind) arg
	logical acc_async_test

Reference:

OpenACC specification v2.6, section 3.2.9.

5.8 acc_async_test_all – Tests for completion of all asynchronous operations.

Description

This function tests for completion of all asynchronous operations. In C/C++, a non-zero value will be returned to indicate all asynchronous operations have completed. While Fortran will return a true. If any asynchronous operation has not completed, C/C++ returns a zero and Fortran returns a false.

C/C++:

Prototype:

int acc_async_test_all(void);

Fortran:

Interface:	function acc_async_test()
	logical acc_get_device_num

Reference:

OpenACC specification v2.6, section 3.2.10.

5.9 acc_wait – Wait for completion of a specific asynchronous operation.

Description

This function waits for completion of the asynchronous operation specified in arg.

C/C++:

Prototype:	acc_wait(arg);
Prototype (OpenACC 1.0 compatibility):	acc_async_wait(arg);

Fortran:

Interface:	subroutine acc_wait(arg)
	integer(acc_handle_kind) arg
Interface (OpenACC 1.0 compatibility):	subroutine acc_async_wait(arg)
	integer(acc_handle_kind) arg

Reference:

OpenACC specification v2.6, section 3.2.11.

5.10 acc_wait_all – Waits for completion of all asynchronous operations.

Description

This function waits for the completion of all asynchronous operations.

C/C++:

Prototype:	acc_wait_all(void);
Prototype (OpenACC 1.0 compatibility):	acc_async_wait_all(void);

Fortran:

Interface:	subroutine acc_wait_all()
Interface (OpenACC 1.0 compatibility):	subroutine acc_async_wait_all()

Reference:

OpenACC specification v2.6, section 3.2.13.

5.11 acc_wait_all_async – Wait for completion of all asynchronous operations.

Description

This function enqueues a wait operation on the queue async for any and all asynchronous operations that have been previously enqueued on any queue.

C/C++:

Prototype:

acc_wait_all_async(int async);

Fortran:

Interface:	subroutine acc_wait_all_async(async)
	integer(acc_handle_kind) async

Reference:

OpenACC specification v2.6, section 3.2.14.

5.12 acc_wait_async – Wait for completion of asynchronous operations.

Description

This function enqueues a wait operation on queue async for any and all asynchronous operations enqueued on queue arg.

C/C++:

Prototype:

acc_wait_async(int arg, int async);

Fortran:

Interface:	subroutine acc_wait_async(arg, async)
	integer(acc_handle_kind) arg, async

Reference:

OpenACC specification v2.6, section 3.2.12.

5.13 acc_init – Initialize runtime for a specific device type.

Description

This function initializes the runtime for the device type specified in devicetype.

C/C++:

Prototype:

acc_init(acc_device_t devicetype);

Fortran:

Interface:	subroutine acc_init(devicetype)
	integer(acc_device_kind) devicetype

Reference:

OpenACC specification v2.6, section 3.2.7.

5.14 acc_shutdown – Shuts down the runtime for a specific device type.

Description

This function shuts down the runtime for the device type specified in devicetype.

C/C++:

Prototype:

acc_shutdown(acc_device_t devicetype);

Fortran:

Interface:	subroutine acc_shutdown(devicetype)
	integer(acc_device_kind) devicetype

Reference:

OpenACC specification v2.6, section 3.2.8.

5.15 acc_on_device – Whether executing on a particular device

Description:

This function returns whether the program is executing on a particular device specified in devicetype. In C/C++ a non-zero value is returned to indicate the device is executing on the specified device type. In Fortran, true will be returned. If the program is not executing on the specified device type C/C++ will return a zero, while Fortran will return false.

C/C++:

Prototype:

acc_on_device(acc_device_t devicetype);

Fortran:

Interface:	function acc_on_device(devicetype)
	integer(acc_device_kind) devicetype
	logical acc_on_device

Reference:

OpenACC specification v2.6, section 3.2.17.

5.16 acc_malloc – Allocate device memory.

Description

This function allocates len bytes of device memory. It returns the device address of the allocated memory.

C/C++:

Prototype:

d_void* acc_malloc(size_t len);

Reference:

OpenACC specification v2.6, section 3.2.18.

5.17 acc_free – Free device memory.

Description

Free previously allocated device memory at the device address a.

C/C++:

Prototype:

acc_free(d_void *a);

Reference:

OpenACC specification v2.6, section 3.2.19.

5.18 acc_copyin – Allocate device memory and copy host memory to it.

Description

In C/C++, this function allocates len bytes of device memory and maps it to the specified host address in a. The device address of the newly allocated device memory is returned.

In Fortran, two (2) forms are supported. In the first form, a specifies a contiguous array section. The second form a specifies a variable or array element and len specifies the length in bytes.

C/C++:

Prototype:	void *acc_copyin(h_void *a, size_t len);
Prototype:	void *acc_copyin_async(h_void *a, size_t len, int async);

Fortran:

Interface:	subroutine acc_copyin(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_copyin(a, len)
	type, dimension(:[,:]...) :: a
	integer len
Interface:	subroutine acc_copyin_async(a, async)
	type, dimension(:[,:]...) :: a
	integer(acc_handle_kind) :: async
Interface:	subroutine acc_copyin_async(a, len, async)
	type, dimension(:[,:]...) :: a
	integer len
	integer(acc_handle_kind) :: async

Reference:

OpenACC specification v2.6, section 3.2.20.

5.19 acc_present_or_copyin – If the data is not present on the device, allocate device memory and copy from host memory.

Description

This function tests if the host data specified by a and of length len is present or not. If it is not present, then device memory will be allocated and the host memory copied. The device address of the newly allocated device memory is returned.

In Fortran, two (2) forms are supported. In the first form, a specifies a contiguous array section. The second form a specifies a variable or array element and len specifies the length in bytes.

Note that acc_present_or_copyin and acc_pcopyin exist for backward compatibility with OpenACC 2.0; use acc_copyin – Allocate device memory and copy host memory to it. instead.

C/C++:

Prototype:	void *acc_present_or_copyin(h_void *a, size_t len);
Prototype:	void *acc_pcopyin(h_void *a, size_t len);

Fortran:

Interface:	subroutine acc_present_or_copyin(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_present_or_copyin(a, len)
	type, dimension(:[,:]...) :: a
	integer len
Interface:	subroutine acc_pcopyin(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_pcopyin(a, len)
	type, dimension(:[,:]...) :: a
	integer len

Reference:

OpenACC specification v2.6, section 3.2.20.

5.20 acc_create – Allocate device memory and map it to host memory.

Description

This function allocates device memory and maps it to host memory specified by the host address a with a length of len bytes. In C/C++, the function returns the device address of the allocated device memory.

In Fortran, two (2) forms are supported. In the first form, a specifies a contiguous array section. The second form a specifies a variable or array element and len specifies the length in bytes.

C/C++:

Prototype:	void *acc_create(h_void *a, size_t len);
Prototype:	void *acc_create_async(h_void *a, size_t len, int async);

Fortran:

Interface:	subroutine acc_create(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_create(a, len)
	type, dimension(:[,:]...) :: a
	integer len
Interface:	subroutine acc_create_async(a, async)
	type, dimension(:[,:]...) :: a
	integer(acc_handle_kind) :: async
Interface:	subroutine acc_create_async(a, len, async)
	type, dimension(:[,:]...) :: a
	integer len
	integer(acc_handle_kind) :: async

Reference:

OpenACC specification v2.6, section 3.2.21.

5.21 acc_present_or_create – If the data is not present on the device, allocate device memory and map it to host memory.

Description

This function tests if the host data specified by a and of length len is present or not. If it is not present, then device memory will be allocated and mapped to host memory. In C/C++, the device address of the newly allocated device memory is returned.

In Fortran, two (2) forms are supported. In the first form, a specifies a contiguous array section. The second form a specifies a variable or array element and len specifies the length in bytes.

Note that acc_present_or_create and acc_pcreate exist for backward compatibility with OpenACC 2.0; use acc_create – Allocate device memory and map it to host memory. instead.

C/C++:

Prototype:	void *acc_present_or_create(h_void *a, size_t len)
Prototype:	void *acc_pcreate(h_void *a, size_t len)

Fortran:

Interface:	subroutine acc_present_or_create(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_present_or_create(a, len)
	type, dimension(:[,:]...) :: a
	integer len
Interface:	subroutine acc_pcreate(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_pcreate(a, len)
	type, dimension(:[,:]...) :: a
	integer len

Reference:

OpenACC specification v2.6, section 3.2.21.

5.22 acc_copyout – Copy device memory to host memory.

Description

This function copies mapped device memory to host memory which is specified by host address a for a length len bytes in C/C++.

In Fortran, two (2) forms are supported. In the first form, a specifies a contiguous array section. The second form a specifies a variable or array element and len specifies the length in bytes.

C/C++:

Prototype:	acc_copyout(h_void *a, size_t len);
Prototype:	acc_copyout_async(h_void *a, size_t len, int async);
Prototype:	acc_copyout_finalize(h_void *a, size_t len);
Prototype:	acc_copyout_finalize_async(h_void *a, size_t len, int async);

Fortran:

Interface:	subroutine acc_copyout(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_copyout(a, len)
	type, dimension(:[,:]...) :: a
	integer len
Interface:	subroutine acc_copyout_async(a, async)
	type, dimension(:[,:]...) :: a
	integer(acc_handle_kind) :: async
Interface:	subroutine acc_copyout_async(a, len, async)
	type, dimension(:[,:]...) :: a
	integer len
	integer(acc_handle_kind) :: async
Interface:	subroutine acc_copyout_finalize(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_copyout_finalize(a, len)
	type, dimension(:[,:]...) :: a
	integer len
Interface:	subroutine acc_copyout_finalize_async(a, async)
	type, dimension(:[,:]...) :: a
	integer(acc_handle_kind) :: async
Interface:	subroutine acc_copyout_finalize_async(a, len, async)
	type, dimension(:[,:]...) :: a
	integer len
	integer(acc_handle_kind) :: async

Reference:

OpenACC specification v2.6, section 3.2.22.

5.23 acc_delete – Free device memory.

Description

This function frees previously allocated device memory specified by the device address a and the length of len bytes.

In Fortran, two (2) forms are supported. In the first form, a specifies a contiguous array section. The second form a specifies a variable or array element and len specifies the length in bytes.

C/C++:

Prototype:	acc_delete(h_void *a, size_t len);
Prototype:	acc_delete_async(h_void *a, size_t len, int async);
Prototype:	acc_delete_finalize(h_void *a, size_t len);
Prototype:	acc_delete_finalize_async(h_void *a, size_t len, int async);

Fortran:

Interface:	subroutine acc_delete(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_delete(a, len)
	type, dimension(:[,:]...) :: a
	integer len
Interface:	subroutine acc_delete_async(a, async)
	type, dimension(:[,:]...) :: a
	integer(acc_handle_kind) :: async
Interface:	subroutine acc_delete_async(a, len, async)
	type, dimension(:[,:]...) :: a
	integer len
	integer(acc_handle_kind) :: async
Interface:	subroutine acc_delete_finalize(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_delete_finalize(a, len)
	type, dimension(:[,:]...) :: a
	integer len
Interface:	subroutine acc_delete_async_finalize(a, async)
	type, dimension(:[,:]...) :: a
	integer(acc_handle_kind) :: async
Interface:	subroutine acc_delete_async_finalize(a, len, async)
	type, dimension(:[,:]...) :: a
	integer len
	integer(acc_handle_kind) :: async

Reference:

OpenACC specification v2.6, section 3.2.23.

5.24 acc_update_device – Update device memory from mapped host memory.

Description

This function updates the device copy from the previously mapped host memory. The host memory is specified with the host address a and a length of len bytes.

In Fortran, two (2) forms are supported. In the first form, a specifies a contiguous array section. The second form a specifies a variable or array element and len specifies the length in bytes.

C/C++:

Prototype:	acc_update_device(h_void *a, size_t len);
Prototype:	acc_update_device(h_void *a, size_t len, async);

Fortran:

Interface:	subroutine acc_update_device(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_update_device(a, len)
	type, dimension(:[,:]...) :: a
	integer len
Interface:	subroutine acc_update_device_async(a, async)
	type, dimension(:[,:]...) :: a
	integer(acc_handle_kind) :: async
Interface:	subroutine acc_update_device_async(a, len, async)
	type, dimension(:[,:]...) :: a
	integer len
	integer(acc_handle_kind) :: async

Reference:

OpenACC specification v2.6, section 3.2.24.

5.25 acc_update_self – Update host memory from mapped device memory.

Description

This function updates the host copy from the previously mapped device memory. The host memory is specified with the host address a and a length of len bytes.

In Fortran, two (2) forms are supported. In the first form, a specifies a contiguous array section. The second form a specifies a variable or array element and len specifies the length in bytes.

C/C++:

Prototype:	acc_update_self(h_void *a, size_t len);
Prototype:	acc_update_self_async(h_void *a, size_t len, int async);

Fortran:

Interface:	subroutine acc_update_self(a)
	type, dimension(:[,:]...) :: a
Interface:	subroutine acc_update_self(a, len)
	type, dimension(:[,:]...) :: a
	integer len
Interface:	subroutine acc_update_self_async(a, async)
	type, dimension(:[,:]...) :: a
	integer(acc_handle_kind) :: async
Interface:	subroutine acc_update_self_async(a, len, async)
	type, dimension(:[,:]...) :: a
	integer len
	integer(acc_handle_kind) :: async

Reference:

OpenACC specification v2.6, section 3.2.25.

5.26 acc_map_data – Map previously allocated device memory to host memory.

Description

This function maps previously allocated device and host memory. The device memory is specified with the device address d. The host memory is specified with the host address h and a length of len.

C/C++:

Prototype:

acc_map_data(h_void *h, d_void *d, size_t len);

Reference:

OpenACC specification v2.6, section 3.2.26.

5.27 acc_unmap_data – Unmap device memory from host memory.

Description

This function unmaps previously mapped device and host memory. The latter specified by h.

C/C++:

Prototype:

acc_unmap_data(h_void *h);

Reference:

OpenACC specification v2.6, section 3.2.27.

5.28 acc_deviceptr – Get device pointer associated with specific host address.

Description

This function returns the device address that has been mapped to the host address specified by h.

C/C++:

Prototype:

void *acc_deviceptr(h_void *h);

Reference:

OpenACC specification v2.6, section 3.2.28.

5.29 acc_hostptr – Get host pointer associated with specific device address.

Description

This function returns the host address that has been mapped to the device address specified by d.

C/C++:

Prototype:

void *acc_hostptr(d_void *d);

Reference:

OpenACC specification v2.6, section 3.2.29.

5.30 acc_is_present – Indicate whether host variable / array is present on device.

Description

This function indicates whether the specified host address in a and a length of len bytes is present on the device. In C/C++, a non-zero value is returned to indicate the presence of the mapped memory on the device. A zero is returned to indicate the memory is not mapped on the device.

In Fortran, two (2) forms are supported. In the first form, a specifies a contiguous array section. The second form a specifies a variable or array element and len specifies the length in bytes. If the host memory is mapped to device memory, then a true is returned. Otherwise, a false is return to indicate the mapped memory is not present.

C/C++:

Prototype:

int acc_is_present(h_void *a, size_t len);

Fortran:

Interface:	function acc_is_present(a)
	type, dimension(:[,:]...) :: a
	logical acc_is_present
Interface:	function acc_is_present(a, len)
	type, dimension(:[,:]...) :: a
	integer len
	logical acc_is_present

Reference:

OpenACC specification v2.6, section 3.2.30.

5.31 acc_memcpy_to_device – Copy host memory to device memory.

Description

This function copies host memory specified by host address of src to device memory specified by the device address dest for a length of bytes bytes.

C/C++:

Prototype:

acc_memcpy_to_device(d_void *dest, h_void *src, size_t bytes);

Reference:

OpenACC specification v2.6, section 3.2.31.

5.32 acc_memcpy_from_device – Copy device memory to host memory.

Description

This function copies host memory specified by host address of src from device memory specified by the device address dest for a length of bytes bytes.

C/C++:

Prototype:

acc_memcpy_from_device(d_void *dest, h_void *src, size_t bytes);

Reference:

OpenACC specification v2.6, section 3.2.32.

5.33 acc_attach – Let device pointer point to device-pointer target.

Description

This function updates a pointer on the device from pointing to a host-pointer address to pointing to the corresponding device data.

C/C++:

Prototype:	acc_attach(h_void **ptr);
Prototype:	acc_attach_async(h_void **ptr, int async);

Reference:

OpenACC specification v2.6, section 3.2.34.

5.34 acc_detach – Let device pointer point to host-pointer target.

Description

This function updates a pointer on the device from pointing to a device-pointer address to pointing to the corresponding host data.

C/C++:

Prototype:	acc_detach(h_void **ptr);
Prototype:	acc_detach_async(h_void **ptr, int async);
Prototype:	acc_detach_finalize(h_void **ptr);
Prototype:	acc_detach_finalize_async(h_void **ptr, int async);

Reference:

OpenACC specification v2.6, section 3.2.35.

5.35 acc_get_current_cuda_device – Get CUDA device handle.

Description

This function returns the CUDA device handle. This handle is the same as used by the CUDA Runtime or Driver API’s.

C/C++:

Prototype:

void *acc_get_current_cuda_device(void);

Reference:

OpenACC specification v2.6, section A.2.1.1.

5.36 acc_get_current_cuda_context – Get CUDA context handle.

Description

This function returns the CUDA context handle. This handle is the same as used by the CUDA Runtime or Driver API’s.

C/C++:

Prototype:

void *acc_get_current_cuda_context(void);

Reference:

OpenACC specification v2.6, section A.2.1.2.

5.37 acc_get_cuda_stream – Get CUDA stream handle.

Description

This function returns the CUDA stream handle for the queue async. This handle is the same as used by the CUDA Runtime or Driver API’s.

C/C++:

Prototype:

void *acc_get_cuda_stream(int async);

Reference:

OpenACC specification v2.6, section A.2.1.3.

5.38 acc_set_cuda_stream – Set CUDA stream handle.

Description

This function associates the stream handle specified by stream with the queue async.

This cannot be used to change the stream handle associated with acc_async_sync.

The return value is not specified.

C/C++:

Prototype:

int acc_set_cuda_stream(int async, void *stream);

Reference:

OpenACC specification v2.6, section A.2.1.4.

5.39 acc_prof_register – Register callbacks.

Description:

This function registers callbacks.

C/C++:

Prototype:

void acc_prof_register (acc_event_t, acc_prof_callback, acc_register_t);

See also:

OpenACC Profiling Interface

Reference:

OpenACC specification v2.6, section 5.3.

5.40 acc_prof_unregister – Unregister callbacks.

Description:

This function unregisters callbacks.

C/C++:

Prototype:

void acc_prof_unregister (acc_event_t, acc_prof_callback, acc_register_t);

See also:

OpenACC Profiling Interface

Reference:

OpenACC specification v2.6, section 5.3.

5.41 acc_prof_lookup – Obtain inquiry functions.

Description:

Function to obtain inquiry functions.

C/C++:

Prototype:

acc_query_fn acc_prof_lookup (const char *);

See also:

OpenACC Profiling Interface

Reference:

OpenACC specification v2.6, section 5.3.

5.42 acc_register_library – Library registration.

Description:

Function for library registration.

C/C++:

Prototype:

void acc_register_library (acc_prof_reg, acc_prof_reg, acc_prof_lookup_func);

See also:

OpenACC Profiling Interface, ACC_PROFLIB

Reference:

OpenACC specification v2.6, section 5.3.

6.1 ACC_DEVICE_TYPE

Reference:

OpenACC specification v2.6, section 4.1.

6.2 ACC_DEVICE_NUM

Reference:

OpenACC specification v2.6, section 4.2.

6.3 ACC_PROFLIB

See also:

acc_register_library – Library registration., OpenACC Profiling Interface

Reference:

OpenACC specification v2.6, section 4.3.

6.4 GCC_ACC_NOTIFY

Description:

Print debug information pertaining to the accelerator.

8.1 Introduction

The OpenACC library uses the CUDA Driver API, and may interact with programs that use the Runtime library directly, or another library based on the Runtime library, e.g., CUBLAS². This chapter describes the use cases and what changes are required in order to use both the OpenACC library and the CUBLAS and Runtime libraries within a program.

8.2 First invocation: NVIDIA CUBLAS library API

In this first use case (see below), a function in the CUBLAS library is called prior to any of the functions in the OpenACC library. More specifically, the function cublasCreate().

When invoked, the function initializes the library and allocates the hardware resources on the host and the device on behalf of the caller. Once the initialization and allocation has completed, a handle is returned to the caller. The OpenACC library also requires initialization and allocation of hardware resources. Since the CUBLAS library has already allocated the hardware resources for the device, all that is left to do is to initialize the OpenACC library and acquire the hardware resources on the host.

Prior to calling the OpenACC function that initializes the library and allocate the host hardware resources, you need to acquire the device number that was allocated during the call to cublasCreate(). The invoking of the runtime library function cudaGetDevice() accomplishes this. Once acquired, the device number is passed along with the device type as parameters to the OpenACC library function acc_set_device_num().

Once the call to acc_set_device_num() has completed, the OpenACC library uses the context that was created during the call to cublasCreate(). In other words, both libraries will be sharing the same context.

    /* Create the handle */
    s = cublasCreate(&h);
    if (s != CUBLAS_STATUS_SUCCESS)
    {
        fprintf(stderr, "cublasCreate failed %d\n", s);
        exit(EXIT_FAILURE);
    }

    /* Get the device number */
    e = cudaGetDevice(&dev);
    if (e != cudaSuccess)
    {
        fprintf(stderr, "cudaGetDevice failed %d\n", e);
        exit(EXIT_FAILURE);
    }

    /* Initialize OpenACC library and use device 'dev' */
    acc_set_device_num(dev, acc_device_nvidia);

8.3 First invocation: OpenACC library API

In this second use case (see below), a function in the OpenACC library is called prior to any of the functions in the CUBLAS library. More specificially, the function acc_set_device_num().

In the use case presented here, the function acc_set_device_num() is used to both initialize the OpenACC library and allocate the hardware resources on the host and the device. In the call to the function, the call parameters specify which device to use and what device type to use, i.e., acc_device_nvidia. It should be noted that this is but one method to initialize the OpenACC library and allocate the appropriate hardware resources. Other methods are available through the use of environment variables and these will be discussed in the next section.

Once the call to acc_set_device_num() has completed, other OpenACC functions can be called as seen with multiple calls being made to acc_copyin(). In addition, calls can be made to functions in the CUBLAS library. In the use case a call to cublasCreate() is made subsequent to the calls to acc_copyin(). As seen in the previous use case, a call to cublasCreate() initializes the CUBLAS library and allocates the hardware resources on the host and the device. However, since the device has already been allocated, cublasCreate() will only initialize the CUBLAS library and allocate the appropriate hardware resources on the host. The context that was created as part of the OpenACC initialization is shared with the CUBLAS library, similarly to the first use case.

    dev = 0;

    acc_set_device_num(dev, acc_device_nvidia);

    /* Copy the first set to the device */
    d_X = acc_copyin(&h_X[0], N * sizeof (float));
    if (d_X == NULL)
    { 
        fprintf(stderr, "copyin error h_X\n");
        exit(EXIT_FAILURE);
    }

    /* Copy the second set to the device */
    d_Y = acc_copyin(&h_Y1[0], N * sizeof (float));
    if (d_Y == NULL)
    { 
        fprintf(stderr, "copyin error h_Y1\n");
        exit(EXIT_FAILURE);
    }

    /* Create the handle */
    s = cublasCreate(&h);
    if (s != CUBLAS_STATUS_SUCCESS)
    {
        fprintf(stderr, "cublasCreate failed %d\n", s);
        exit(EXIT_FAILURE);
    }

    /* Perform saxpy using CUBLAS library function */
    s = cublasSaxpy(h, N, &alpha, d_X, 1, d_Y, 1);
    if (s != CUBLAS_STATUS_SUCCESS)
    {
        fprintf(stderr, "cublasSaxpy failed %d\n", s);
        exit(EXIT_FAILURE);
    }

    /* Copy the results from the device */
    acc_memcpy_from_device(&h_Y1[0], d_Y, N * sizeof (float));

8.4 OpenACC library and environment variables

There are two environment variables associated with the OpenACC library that may be used to control the device type and device number: ACC_DEVICE_TYPE and ACC_DEVICE_NUM, respectively. These two environment variables can be used as an alternative to calling acc_set_device_num(). As seen in the second use case, the device type and device number were specified using acc_set_device_num(). If however, the aforementioned environment variables were set, then the call to acc_set_device_num() would not be required.

The use of the environment variables is only relevant when an OpenACC function is called prior to a call to cudaCreate(). If cudaCreate() is called prior to a call to an OpenACC function, then you must call acc_set_device_num()³

9.1 Implementation Status and Implementation-Defined Behavior

We’re implementing the OpenACC Profiling Interface as defined by the OpenACC 2.6 specification. We’re clarifying some aspects here as implementation-defined behavior, while they’re still under discussion within the OpenACC Technical Committee.

This implementation is tuned to keep the performance impact as low as possible for the (very common) case that the Profiling Interface is not enabled. This is relevant, as the Profiling Interface affects all the hot code paths (in the target code, not in the offloaded code). Users of the OpenACC Profiling Interface can be expected to understand that performance will be impacted to some degree once the Profiling Interface has gotten enabled: for example, because of the runtime (libgomp) calling into a third-party library for every event that has been registered.

We’re not yet accounting for the fact that OpenACC events may occur during event processing. We just handle one case specially, as required by CUDA 9.0 nvprof, that acc_get_device_type (acc_get_device_type – Get type of device accelerator to be used.)) may be called from acc_ev_device_init_start, acc_ev_device_init_end callbacks.

We’re not yet implementing initialization via a acc_register_library function that is either statically linked in, or dynamically via LD_PRELOAD. Initialization via acc_register_library functions dynamically loaded via the ACC_PROFLIB environment variable does work, as does directly calling acc_prof_register, acc_prof_unregister, acc_prof_lookup.

As currently there are no inquiry functions defined, calls to acc_prof_lookup will always return NULL.

There aren’t separate start, stop events defined for the event types acc_ev_create, acc_ev_delete, acc_ev_alloc, acc_ev_free. It’s not clear if these should be triggered before or after the actual device-specific call is made. We trigger them after.

Remarks about data provided to callbacks:

acc_prof_info.event_type

It’s not clear if for nested event callbacks (for example, acc_ev_enqueue_launch_start as part of a parent compute construct), this should be set for the nested event (acc_ev_enqueue_launch_start), or if the value of the parent construct should remain (acc_ev_compute_construct_start). In this implementation, the value will generally correspond to the innermost nested event type.

acc_prof_info.device_type

• For acc_ev_compute_construct_start, and in presence of an if clause with false argument, this will still refer to the offloading device type. It’s not clear if that’s the expected behavior.
• Complementary to the item before, for acc_ev_compute_construct_end, this is set to acc_device_host in presence of an if clause with false argument. It’s not clear if that’s the expected behavior.

acc_prof_info.thread_id

Always -1; not yet implemented.

acc_prof_info.async

• Not yet implemented correctly for acc_ev_compute_construct_start.
• In a compute construct, for host-fallback execution/acc_device_host it will always be acc_async_sync. It’s not clear if that’s the expected behavior.
• For acc_ev_device_init_start and acc_ev_device_init_end, it will always be acc_async_sync. It’s not clear if that’s the expected behavior.

acc_prof_info.async_queue

There is no limited number of asynchronous queues in libgomp. This will always have the same value as acc_prof_info.async.

acc_prof_info.src_file

Always NULL; not yet implemented.

acc_prof_info.func_name

Always NULL; not yet implemented.

acc_prof_info.line_no

Always -1; not yet implemented.

acc_prof_info.end_line_no

Always -1; not yet implemented.

acc_prof_info.func_line_no

Always -1; not yet implemented.

acc_prof_info.func_end_line_no

Always -1; not yet implemented.

acc_event_info.event_type, acc_event_info.*.event_type

Relating to acc_prof_info.event_type discussed above, in this implementation, this will always be the same value as acc_prof_info.event_type.

acc_event_info.*.parent_construct

• Will be acc_construct_parallel for all OpenACC compute constructs as well as many OpenACC Runtime API calls; should be the one matching the actual construct, or acc_construct_runtime_api, respectively.
• Will be acc_construct_enter_data or acc_construct_exit_data when processing variable mappings specified in OpenACC declare directives; should be acc_construct_declare.
• For implicit acc_ev_device_init_start, acc_ev_device_init_end, and explicit as well as implicit acc_ev_alloc, acc_ev_free, acc_ev_enqueue_upload_start, acc_ev_enqueue_upload_end, acc_ev_enqueue_download_start, and acc_ev_enqueue_download_end, will be acc_construct_parallel; should reflect the real parent construct.

acc_event_info.*.implicit

For acc_ev_alloc, acc_ev_free, acc_ev_enqueue_upload_start, acc_ev_enqueue_upload_end, acc_ev_enqueue_download_start, and acc_ev_enqueue_download_end, this currently will be 1 also for explicit usage.

acc_event_info.data_event.var_name

Always NULL; not yet implemented.

acc_event_info.data_event.host_ptr

For acc_ev_alloc, and acc_ev_free, this is always NULL.

typedef union acc_api_info

… as printed in 5.2.3. Third Argument: API-Specific Information. This should obviously be typedef struct acc_api_info.

acc_api_info.device_api

Possibly not yet implemented correctly for acc_ev_compute_construct_start, acc_ev_device_init_start, acc_ev_device_init_end: will always be acc_device_api_none for these event types. For acc_ev_enter_data_start, it will be acc_device_api_none in some cases.

acc_api_info.device_type

Always the same as acc_prof_info.device_type.

acc_api_info.vendor

Always -1; not yet implemented.

acc_api_info.device_handle

Always NULL; not yet implemented.

acc_api_info.context_handle

Always NULL; not yet implemented.

acc_api_info.async_handle

Always NULL; not yet implemented.

Remarks about certain event types:

acc_ev_device_init_start, acc_ev_device_init_end

• Whan a compute construct triggers implicit acc_ev_device_init_start and acc_ev_device_init_end events, they currently aren’t nested within the corresponding acc_ev_compute_construct_start and acc_ev_compute_construct_end, but they’re currently observed before acc_ev_compute_construct_start. It’s not clear what to do: the standard asks us provide a lot of details to the acc_ev_compute_construct_start callback, without (implicitly) initializing a device before?
• Callbacks for these event types will not be invoked for calls to the acc_set_device_type and acc_set_device_num functions. It’s not clear if they should be.

acc_ev_enter_data_start, acc_ev_enter_data_end, acc_ev_exit_data_start, acc_ev_exit_data_end

• Callbacks for these event types will also be invoked for OpenACC host_data constructs. It’s not clear if they should be.
• Callbacks for these event types will also be invoked when processing variable mappings specified in OpenACC declare directives. It’s not clear if they should be.

Callbacks for the following event types will be invoked, but dispatch and information provided therein has not yet been thoroughly reviewed:

• acc_ev_alloc
• acc_ev_free
• acc_ev_update_start, acc_ev_update_end
• acc_ev_enqueue_upload_start, acc_ev_enqueue_upload_end
• acc_ev_enqueue_download_start, acc_ev_enqueue_download_end

During device initialization, and finalization, respectively, callbacks for the following event types will not yet be invoked:

• acc_ev_alloc
• acc_ev_free

Callbacks for the following event types have not yet been implemented, so currently won’t be invoked:

• acc_ev_device_shutdown_start, acc_ev_device_shutdown_end
• acc_ev_runtime_shutdown
• acc_ev_create, acc_ev_delete
• acc_ev_wait_start, acc_ev_wait_end

For the following runtime library functions, not all expected callbacks will be invoked (mostly concerning implicit device initialization):

• acc_get_num_devices
• acc_set_device_type
• acc_get_device_type
• acc_set_device_num
• acc_get_device_num
• acc_init
• acc_shutdown

Aside from implicit device initialization, for the following runtime library functions, no callbacks will be invoked for shared-memory offloading devices (it’s not clear if they should be):

• acc_malloc
• acc_free
• acc_copyin, acc_present_or_copyin, acc_copyin_async
• acc_create, acc_present_or_create, acc_create_async
• acc_copyout, acc_copyout_async, acc_copyout_finalize, acc_copyout_finalize_async
• acc_delete, acc_delete_async, acc_delete_finalize, acc_delete_finalize_async
• acc_update_device, acc_update_device_async
• acc_update_self, acc_update_self_async
• acc_map_data, acc_unmap_data
• acc_memcpy_to_device, acc_memcpy_to_device_async
• acc_memcpy_from_device, acc_memcpy_from_device_async

10.1 Implementing MASTER construct

if (omp_get_thread_num () == 0)
  block

Alternately, we generate two copies of the parallel subfunction and only include this in the version run by the master thread. Surely this is not worthwhile though...

10.2 Implementing CRITICAL construct

Without a specified name,

  void GOMP_critical_start (void);
  void GOMP_critical_end (void);

so that we don’t get COPY relocations from libgomp to the main application.

With a specified name, use omp_set_lock and omp_unset_lock with name being transformed into a variable declared like

  omp_lock_t gomp_critical_user_<name> __attribute__((common))

Ideally the ABI would specify that all zero is a valid unlocked state, and so we wouldn’t need to initialize this at startup.

10.3 Implementing ATOMIC construct

The target should implement the __sync builtins.

Failing that we could add

  void GOMP_atomic_enter (void)
  void GOMP_atomic_exit (void)

which reuses the regular lock code, but with yet another lock object private to the library.

10.4 Implementing FLUSH construct

Expands to the __sync_synchronize builtin.

10.5 Implementing BARRIER construct

  void GOMP_barrier (void)

10.6 Implementing THREADPRIVATE construct

In _most_ cases we can map this directly to __thread. Except that OMP allows constructors for C++ objects. We can either refuse to support this (how often is it used?) or we can implement something akin to .ctors.

Even more ideally, this ctor feature is handled by extensions to the main pthreads library. Failing that, we can have a set of entry points to register ctor functions to be called.

10.7 Implementing PRIVATE clause

In association with a PARALLEL, or within the lexical extent of a PARALLEL block, the variable becomes a local variable in the parallel subfunction.

In association with FOR or SECTIONS blocks, create a new automatic variable within the current function. This preserves the semantic of new variable creation.

10.8 Implementing FIRSTPRIVATE LASTPRIVATE COPYIN and COPYPRIVATE clauses

This seems simple enough for PARALLEL blocks. Create a private struct for communicating between the parent and subfunction. In the parent, copy in values for scalar and "small" structs; copy in addresses for others TREE_ADDRESSABLE types. In the subfunction, copy the value into the local variable.

It is not clear what to do with bare FOR or SECTION blocks. The only thing I can figure is that we do something like:

#pragma omp for firstprivate(x) lastprivate(y)
for (int i = 0; i < n; ++i)
  body;

which becomes

{
  int x = x, y;

  // for stuff

  if (i == n)
    y = y;
}

where the "x=x" and "y=y" assignments actually have different uids for the two variables, i.e. not something you could write directly in C. Presumably this only makes sense if the "outer" x and y are global variables.

COPYPRIVATE would work the same way, except the structure broadcast would have to happen via SINGLE machinery instead.

10.9 Implementing REDUCTION clause

The private struct mentioned in the previous section should have a pointer to an array of the type of the variable, indexed by the thread’s team_id. The thread stores its final value into the array, and after the barrier, the master thread iterates over the array to collect the values.

10.10 Implementing PARALLEL construct

  #pragma omp parallel
  {
    body;
  }

becomes

  void subfunction (void *data)
  {
    use data;
    body;
  }

  setup data;
  GOMP_parallel_start (subfunction, &data, num_threads);
  subfunction (&data);
  GOMP_parallel_end ();

  void GOMP_parallel_start (void (*fn)(void *), void *data, unsigned num_threads)

The FN argument is the subfunction to be run in parallel.

The DATA argument is a pointer to a structure used to communicate data in and out of the subfunction, as discussed above with respect to FIRSTPRIVATE et al.

The NUM_THREADS argument is 1 if an IF clause is present and false, or the value of the NUM_THREADS clause, if present, or 0.

The function needs to create the appropriate number of threads and/or launch them from the dock. It needs to create the team structure and assign team ids.

  void GOMP_parallel_end (void)

Tears down the team and returns us to the previous omp_in_parallel() state.

10.11 Implementing FOR construct

  #pragma omp parallel for
  for (i = lb; i <= ub; i++)
    body;

becomes

  void subfunction (void *data)
  {
    long _s0, _e0;
    while (GOMP_loop_static_next (&_s0, &_e0))
    {
      long _e1 = _e0, i;
      for (i = _s0; i < _e1; i++)
        body;
    }
    GOMP_loop_end_nowait ();
  }

  GOMP_parallel_loop_static (subfunction, NULL, 0, lb, ub+1, 1, 0);
  subfunction (NULL);
  GOMP_parallel_end ();

  #pragma omp for schedule(runtime)
  for (i = 0; i < n; i++)
    body;

becomes

  {
    long i, _s0, _e0;
    if (GOMP_loop_runtime_start (0, n, 1, &_s0, &_e0))
      do {
        long _e1 = _e0;
        for (i = _s0, i < _e0; i++)
          body;
      } while (GOMP_loop_runtime_next (&_s0, _&e0));
    GOMP_loop_end ();
  }

Note that while it looks like there is trickiness to propagating a non-constant STEP, there isn’t really. We’re explicitly allowed to evaluate it as many times as we want, and any variables involved should automatically be handled as PRIVATE or SHARED like any other variables. So the expression should remain evaluable in the subfunction. We can also pull it into a local variable if we like, but since its supposed to remain unchanged, we can also not if we like.

If we have SCHEDULE(STATIC), and no ORDERED, then we ought to be able to get away with no work-sharing context at all, since we can simply perform the arithmetic directly in each thread to divide up the iterations. Which would mean that we wouldn’t need to call any of these routines.

There are separate routines for handling loops with an ORDERED clause. Bookkeeping for that is non-trivial...

10.12 Implementing ORDERED construct

  void GOMP_ordered_start (void)
  void GOMP_ordered_end (void)

10.13 Implementing SECTIONS construct

A block as

  #pragma omp sections
  {
    #pragma omp section
    stmt1;
    #pragma omp section
    stmt2;
    #pragma omp section
    stmt3;
  }

becomes

  for (i = GOMP_sections_start (3); i != 0; i = GOMP_sections_next ())
    switch (i)
      {
      case 1:
        stmt1;
        break;
      case 2:
        stmt2;
        break;
      case 3:
        stmt3;
        break;
      }
  GOMP_barrier ();

10.14 Implementing SINGLE construct

A block like

  #pragma omp single
  {
    body;
  }

becomes

  if (GOMP_single_start ())
    body;
  GOMP_barrier ();

while

  #pragma omp single copyprivate(x)
    body;

becomes

  datap = GOMP_single_copy_start ();
  if (datap == NULL)
    {
      body;
      data.x = x;
      GOMP_single_copy_end (&data);
    }
  else
    x = datap->x;
  GOMP_barrier ();

10.15 Implementing OpenACC’s PARALLEL construct

  void GOACC_parallel ()