starpu-max/doc/doxygen/chapters/480_openmp_runtime_support.doxy

/* StarPU --- Runtime system for heterogeneous multicore architectures.
 *
 * Copyright (C) 2014-2021  Université de Bordeaux, CNRS (LaBRI UMR 5800), Inria
 *
 * StarPU is free software; you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published by
 * the Free Software Foundation; either version 2.1 of the License, or (at
 * your option) any later version.
 *
 * StarPU is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 *
 * See the GNU Lesser General Public License in COPYING.LGPL for more details.
 */

/*! \page OpenMPRuntimeSupport The StarPU OpenMP Runtime Support (SORS)

StarPU provides the necessary routines and support to implement an OpenMP
(http://www.openmp.org/) runtime compliant with the
revision 3.1 of the language specification, and compliant with the
task-related data dependency functionalities introduced in the revision
4.0 of the language. This StarPU OpenMP Runtime Support (SORS) has been
designed to be targetted by OpenMP compilers such as the Klang-OMP
compiler. Most supported OpenMP directives can both be implemented
inline or as outlined functions.

All functions are defined in \ref API_OpenMP_Runtime_Support.

\section OMPImplementation Implementation Details and Specificities

\subsection OMPMainThread Main Thread

When using the SORS, the main thread gets involved in executing OpenMP tasks
just like every other threads, in order to be compliant with the
specification execution model. This contrasts with StarPU's usual
execution model where the main thread submit tasks but does not take
part in executing them.

\subsection OMPTaskSemantics Extended Task Semantics

The semantics of tasks generated by the SORS are extended with respect
to regular StarPU tasks in that SORS' tasks may block and be preempted
by SORS call, whereas regular StarPU tasks cannot. SORS tasks may
coexist with regular StarPU tasks. However, only the tasks created using
SORS API functions inherit from extended semantics.

\section OMPConfiguration Configuration

The SORS can be compiled into <c>libstarpu</c> through
the \c configure option \ref enable-openmp "--enable-openmp".
Conditional compiled source codes may check for the
availability of the OpenMP Runtime Support by testing whether the C
preprocessor macro <c>STARPU_OPENMP</c> is defined or not.

\section OMPInitExit Initialization and Shutdown

The SORS needs to be executed/terminated by the
starpu_omp_init() / starpu_omp_shutdown() instead of
starpu_init() / starpu_shutdown(). This requirement is necessary to make
sure that the main thread gets the proper execution environment to run
OpenMP tasks. These calls will usually be performed by a compiler
runtime. Thus, they can be executed from a constructor/destructor such
as this:

\code{.c}
__attribute__((constructor))
static void omp_constructor(void)
{
	int ret = starpu_omp_init();
	STARPU_CHECK_RETURN_VALUE(ret, "starpu_omp_init");
}

__attribute__((destructor))
static void omp_destructor(void)
{
	starpu_omp_shutdown();
}
\endcode

\sa starpu_omp_init()
\sa starpu_omp_shutdown()

\section OMPSharing Parallel Regions and Worksharing

The SORS provides functions to create OpenMP parallel regions as well as
mapping work on participating workers. The current implementation does
not provide nested active parallel regions: Parallel regions may be
created recursively, however only the first level parallel region may
have more than one worker. From an internal point-of-view, the SORS'
parallel regions are implemented as a set of implicit, extended semantics
StarPU tasks, following the execution model of the OpenMP specification.
Thus the SORS' parallel region tasks may block and be preempted, by
SORS calls, enabling constructs such as barriers.

\subsection OMPParallel Parallel Regions

Parallel regions can be created with the function
starpu_omp_parallel_region() which accepts a set of attributes as
parameter. The execution of the calling task is suspended until the
parallel region completes. The field starpu_omp_parallel_region_attr::cl
is a regular StarPU codelet. However only CPU codelets are
supported for parallel regions.
Here is an example of use:

\code{.c}
void parallel_region_f(void *buffers[], void *args)
{
	(void) buffers;
	(void) args;
	pthread_t tid = pthread_self();
	int worker_id = starpu_worker_get_id();
	printf("[tid %p] task thread = %d\n", (void *)tid, worker_id);
}

void f(void)
{
	struct starpu_omp_parallel_region_attr attr;
	memset(&attr, 0, sizeof(attr));
	attr.cl.cpu_funcs[0] = parallel_region_f;
	attr.cl.where        = STARPU_CPU;
	attr.if_clause       = 1;
	starpu_omp_parallel_region(&attr);
	return 0;
}
\endcode

\sa struct starpu_omp_parallel_region_attr
\sa starpu_omp_parallel_region()

\subsection OMPFor Parallel For

OpenMP <c>for</c> loops are provided by the starpu_omp_for() group of
functions. Variants are available for inline or outlined
implementations. The SORS supports <c>static</c>, <c>dynamic</c>, and
<c>guided</c> loop scheduling clauses. The <c>auto</c> scheduling clause
is implemented as <c>static</c>. The <c>runtime</c> scheduling clause
honors the scheduling mode selected through the environment variable
\c OMP_SCHEDULE or the starpu_omp_set_schedule() function. For loops with
the <c>ordered</c> clause are also supported. An implicit barrier can be
enforced or skipped at the end of the worksharing construct, according
to the value of the <c>nowait</c> parameter.

The canonical family of starpu_omp_for() functions provide each instance
with the first iteration number and the number of iterations (possibly
zero) to perform. The alternate family of starpu_omp_for_alt() functions
provide each instance with the (possibly empty) range of iterations to
perform, including the first and excluding the last.

The family of starpu_omp_ordered() functions enable to implement
OpenMP's ordered construct, a region with a parallel for loop that is
guaranteed to be executed in the sequential order of the loop
iterations.

\code{.c}
void for_g(unsigned long long i, unsigned long long nb_i, void *arg)
{
	(void) arg;
	for (; nb_i > 0; i++, nb_i--)
	{
		array[i] = 1;
	}
}

void parallel_region_f(void *buffers[], void *args)
{
	(void) buffers;
	(void) args;
	starpu_omp_for(for_g, NULL, NB_ITERS, CHUNK, starpu_omp_sched_static, 0, 0);
}
\endcode

\sa starpu_omp_for()
\sa starpu_omp_for_inline_first()
\sa starpu_omp_for_inline_next()
\sa starpu_omp_for_alt()
\sa starpu_omp_for_inline_first_alt()
\sa starpu_omp_for_inline_next_alt()
\sa starpu_omp_ordered()
\sa starpu_omp_ordered_inline_begin()
\sa starpu_omp_ordered_inline_end()


\subsection OMPSections Sections
OpenMP <c>sections</c> worksharing constructs are supported using the
set of starpu_omp_sections() variants. The general principle is either
to provide an array of per-section functions or a single function that
will redirect to execution to the suitable per-section functions. An
implicit barrier can be enforced or skipped at the end of the
worksharing construct, according to the value of the <c>nowait</c>
parameter.

\code{.c}
void parallel_region_f(void *buffers[], void *args)
{
	(void) buffers;
	(void) args;

	section_funcs[0] = f;
	section_funcs[1] = g;
	section_funcs[2] = h;
	section_funcs[3] = i;

	section_args[0] = arg_f;
	section_args[1] = arg_g;
	section_args[2] = arg_h;
	section_args[3] = arg_i;

	starpu_omp_sections(4, section_f, section_args, 0);
}
\endcode

\sa starpu_omp_sections()
\sa starpu_omp_sections_combined()

\subsection OMPSingle Single
OpenMP <c>single</c> workharing constructs are supported using the set
of starpu_omp_single() variants. An
implicit barrier can be enforced or skipped at the end of the
worksharing construct, according to the value of the <c>nowait</c>
parameter.

\code{.c}
void single_f(void *arg)
{
	(void) arg;
	pthread_t tid = pthread_self();
	int worker_id = starpu_worker_get_id();
	printf("[tid %p] task thread = %d -- single\n", (void *)tid, worker_id);
}

void parallel_region_f(void *buffers[], void *args)
{
	(void) buffers;
	(void) args;
	starpu_omp_single(single_f, NULL, 0);
}
\endcode

The SORS also provides dedicated support for  <c>single</c> sections
with <c>copyprivate</c> clauses through the
starpu_omp_single_copyprivate() function variants. The OpenMP
<c>master</c> directive is supported as well using the
starpu_omp_master() function variants.

\sa starpu_omp_master()
\sa starpu_omp_master_inline()
\sa starpu_omp_single()
\sa starpu_omp_single_inline()
\sa starpu_omp_single_copyprivate()
\sa starpu_omp_single_copyprivate_inline_begin()
\sa starpu_omp_single_copyprivate_inline_end()

\section OMPTask Tasks

The SORS implements the necessary support of OpenMP 3.1 and OpenMP 4.0's
so-called explicit tasks, together with OpenMP 4.0's data dependency
management.

\subsection OMPTaskExplicit Explicit Tasks
Explicit OpenMP tasks are created with the SORS using the
starpu_omp_task_region() function. The implementation supports
<c>if</c>, <c>final</c>, <c>untied</c> and <c>mergeable</c> clauses
as defined in the OpenMP specification. Unless specified otherwise by
the appropriate clause(s), the created task may be executed by any
participating worker of the current parallel region.

The current SORS implementation requires explicit tasks to be created
within the context of an active parallel region. In particular, an
explicit task cannot be created by the main thread outside of a parallel
region. Explicit OpenMP tasks created using starpu_omp_task_region() are
implemented as StarPU tasks with extended semantics, and may as such be
blocked and preempted by SORS routines.

The current SORS implementation supports recursive explicit tasks
creation, to ensure compliance with the OpenMP specification. However,
it should be noted that StarPU is not designed nor optimized for
efficiently scheduling of recursive task applications.

The code below shows how to create 4 explicit tasks within a parallel
region.

\code{.c}
void task_region_g(void *buffers[], void *args)
{
	(void) buffers;
	(void) args;
	pthread tid = pthread_self();
	int worker_id = starpu_worker_get_id();
	printf("[tid %p] task thread = %d: explicit task \"g\"\n", (void *)tid, worker_id);
}

void parallel_region_f(void *buffers[], void *args)
{
	(void) buffers;
	(void) args;
	struct starpu_omp_task_region_attr attr;

	memset(&attr, 0, sizeof(attr));
	attr.cl.cpu_funcs[0]  = task_region_g;
	attr.cl.where         = STARPU_CPU;
	attr.if_clause        = 1;
	attr.final_clause     = 0;
	attr.untied_clause    = 1;
	attr.mergeable_clause = 0;
	starpu_omp_task_region(&attr);
	starpu_omp_task_region(&attr);
	starpu_omp_task_region(&attr);
	starpu_omp_task_region(&attr);
}
\endcode

\sa struct starpu_omp_task_region_attr
\sa starpu_omp_task_region()

\subsection OMPDataDependencies Data Dependencies
The SORS implements inter-tasks data dependencies as specified in OpenMP
4.0. Data dependencies are expressed using regular StarPU data handles
(\ref starpu_data_handle_t) plugged into the task's <c>attr.cl</c>
codelet. The family of starpu_vector_data_register() -like functions and the
starpu_data_lookup() function may be used to register a memory area and
to retrieve the current data handle associated with a pointer
respectively. The testcase <c>./tests/openmp/task_02.c</c> gives a
detailed example of using OpenMP 4.0 tasks dependencies with the SORS
implementation.

Note: the OpenMP 4.0 specification only supports data dependencies
between sibling tasks, that is tasks created by the same implicit or
explicit parent task. The current SORS implementation also only supports data
dependencies between sibling tasks. Consequently the behaviour is
unspecified if dependencies are expressed beween tasks that have not
been created by the same parent task.

\subsection OMPTaskSyncs TaskWait and TaskGroup
The SORS implements both the <c>taskwait</c> and <c>taskgroup</c> OpenMP
task synchronization constructs specified in OpenMP 4.0, with the
starpu_omp_taskwait() and starpu_omp_taskgroup() functions respectively.

An example of starpu_omp_taskwait() use, creating two explicit tasks and
waiting for their completion:

\code{.c}
void task_region_g(void *buffers[], void *args)
{
	(void) buffers;
	(void) args;
	printf("Hello, World!\n");
}

void parallel_region_f(void *buffers[], void *args)
{
	(void) buffers;
	(void) args;
	struct starpu_omp_task_region_attr attr;
	memset(&attr, 0, sizeof(attr));
	attr.cl.cpu_funcs[0]  = task_region_g;
	attr.cl.where         = STARPU_CPU;
	attr.if_clause        = 1;
	attr.final_clause     = 0;
	attr.untied_clause    = 1;
	attr.mergeable_clause = 0;
	starpu_omp_task_region(&attr);
	starpu_omp_task_region(&attr);
	starpu_omp_taskwait();
\endcode

An example of starpu_omp_taskgroup() use, creating a task group of two explicit tasks:

\code{.c}
void task_region_g(void *buffers[], void *args)
{
	(void) buffers;
	(void) args;
	printf("Hello, World!\n");
}

void taskgroup_f(void *arg)
{
	(void)arg;
	struct starpu_omp_task_region_attr attr;
	memset(&attr, 0, sizeof(attr));
	attr.cl.cpu_funcs[0]  = task_region_g;
	attr.cl.where         = STARPU_CPU;
	attr.if_clause        = 1;
	attr.final_clause     = 0;
	attr.untied_clause    = 1;
	attr.mergeable_clause = 0;
	starpu_omp_task_region(&attr);
	starpu_omp_task_region(&attr);
}

void parallel_region_f(void *buffers[], void *args)
{
	(void) buffers;
	(void) args;
	starpu_omp_taskgroup(taskgroup_f, (void *)NULL);
}
\endcode

\sa starpu_omp_task_region()
\sa starpu_omp_taskwait()
\sa starpu_omp_taskgroup()
\sa starpu_omp_taskgroup_inline_begin()
\sa starpu_omp_taskgroup_inline_end()

\section OMPSynchronization Synchronization Support

The SORS implements objects and method to build common OpenMP
synchronization constructs.

\subsection OMPSimpleLock Simple Locks

The SORS Simple Locks are opaque starpu_omp_lock_t objects enabling multiple
tasks to synchronize with each others, following the Simple Lock
constructs defined by the OpenMP specification. In accordance with such
specification, simple locks may not by acquired multiple times by the
same task, without being released in-between; otherwise, deadlocks may
result. Codes requiring the possibility to lock multiple times
recursively should use Nestable Locks (\ref NestableLock). Codes NOT
requiring the possibility to lock multiple times recursively should use
Simple Locks as they incur less processing overhead than Nestable Locks.

\sa starpu_omp_lock_t
\sa starpu_omp_init_lock()
\sa starpu_omp_destroy_lock()
\sa starpu_omp_set_lock()
\sa starpu_omp_unset_lock()
\sa starpu_omp_test_lock()

\subsection OMPNestableLock Nestable Locks

The SORS Nestable Locks are opaque starpu_omp_nest_lock_t objects enabling
multiple tasks to synchronize with each others, following the Nestable
Lock constructs defined by the OpenMP specification. In accordance with
such specification, nestable locks may by acquired multiple times
recursively by the same task without deadlocking. Nested locking and
unlocking operations must be well parenthesized at any time, otherwise
deadlock and/or undefined behaviour may occur.  Codes requiring the
possibility to lock multiple times recursively should use Nestable
Locks. Codes NOT requiring the possibility to lock multiple times
recursively should use Simple Locks (\ref SimpleLock) instead, as they
incur less processing overhead than Nestable Locks.

\sa starpu_omp_nest_lock_t
\sa starpu_omp_init_nest_lock()
\sa starpu_omp_destroy_nest_lock()
\sa starpu_omp_set_nest_lock()
\sa starpu_omp_unset_nest_lock()
\sa starpu_omp_test_nest_lock()

\subsection OMPCritical Critical Sections

The SORS implements support for OpenMP critical sections through the
family of \ref starpu_omp_critical functions. Critical sections may optionally
be named. There is a single, common anonymous critical section. Mutual
exclusion only occur within the scope of single critical section, either
a named one or the anonymous one.

\sa starpu_omp_critical()
\sa starpu_omp_critical_inline_begin()
\sa starpu_omp_critical_inline_end()

\subsection OMPBarrier Barriers

The SORS provides the starpu_omp_barrier() function to implement
barriers over parallel region teams. In accordance with the OpenMP
specification, the starpu_omp_barrier() function waits for every
implicit task of the parallel region to reach the barrier and every
explicit task launched by the parallel region to complete, before
returning.

\sa starpu_omp_barrier()

*/