OpenMP 3.0 Implementation-Defined BehaviorsΒΆ

The OpenMP 3.0 specification identifies some features as implementation defined. The following documents the behaviour of the runtime with respect to implementation-defined features.

  • Nested Parallelism: Only one level of parallelism is supported. Nested parallel regions will be executed by teams comprising only one thread.
  • Task Scheduling Points in Untied Tasks: Only tied tasks are currently supported and untied tasks are treated as tied tasks. Therefore, task scheduling points in untied task regions occur at the same points as in tied task regions.
  • Memory Model: Memory accesses by multiple threads to the same variable without synchronization may not be atomic with respect to each other.
  • Dynamic adjustment of threads: Dynamic adjustment of the number of threads is not supported. When a thread encounters a parallel construct, the number of threads delivered by this implementation is determined according to Algorithm 2.1 in the OpenMP 3.0 Specification. When more threads are requested than are available, the available threads are supplied.
  • Loop directive: The effect of schedule(runtime) clause when the run-sched-var ICV is set to auto is static with no chunk size.
  • Constructs
    • sections construct: The structured blocks in a sections construct are assigned to the members of the team executing the sections region, such that the threads execute an approximately equal number of sections.
    • single construct: The first thread that encounters a single construct executes its structured block.
  • Runtime library routines
    • omp_set_num_theads: If the argument is not a positive integer, then the call is ignored.
    • omp_set_schedule: Only valid OpenMP schedule kinds are accepted.
    • omp_set_max_active_levels: When called from within an explicit parallel region, the binding thread set is all threads. However, since nested parallelism is not supported, this implementation has little use for this function. If the argument is a negative integer or a positive integer greater than one, then the call is ignored.
    • omp_get_max_active_levels: This routine may be called anywhere in the program. It always returns the value of the gomp_max_active_levels_var internal control variable.