To: J3                                                     J3/23-227
From: Bill Long
Subject:  OpenMP report
Date: 2023-October-16

Since the last report, the OpenMP language committee met once more to
prepare the upcoming technical report (TR12), to be released at
SC23. There will be more more TR released next summer, a 6.0 comment
draft, before the final OpenMP 6.0 document is released (expected at
SC24).



The following are a list of new features that have been voted in:



* Fixes to the specification of the new coexecute construct, enabling
  array syntax and intrinsic functions to parallelized more
  effectively in target regions.



* Permit the loop directive (and combined constructs that end in the
  loop directive) to be applied to a DO CONCURRENT loop. This gives
  more control to the user on how a DO CONCURRENT loop should utilize
  the parallelism available in the OpenMP program. For instance, to
  offload a DO CONCURRENT loop to an accelerator, one can write:



   !$omp target teams loop

   do concurrent (i=1:n,j=1:n)

      ...

   end do



There are limitations on the clauses that can appear on the loop
directive, to avoid clashes with the DO CONCURRENT
semantic. Specifically, no collapse clause and no data-sharing
attribute clauses may be specified. OpenMP also restricts applying a
directive to a DO CONCURRENT loop if there are conflicting accesses to
a variable in multiple iterations (at least is a modification) unless
a locality is explicitly specified for the variable.



* Support for "free-agent threads" was added. This allows certain
  tasks to be scheduled on an available thread in an implementation's
  thread pool that is not part of the current thread team. This allows
  the program to achieve additional concurrency when generating tasks
  even if an active parallel region is not present.



* The 'fuse' loop transformation construct was added, enabling a
  sequence of consecutive loop nests to be fused. Prior loop
  transformation constructs we have added include: unroll, tile,
  interchange, and reverse.



* A 'self' modifier was added to the 'map' clause. This directs the
  implementation to map the list item without greating a new copy in
  device memory (i.e., the address of the list item is mapped to
  itself). While this would mostly commonly be used on systems that
  support unified shared memory, it could be supported on other
  systems as well through address pinning.



* Add omp_target_memset and omp_target_memset_sync routines.

 These are some of the other big topics we are looking to fix or
 address before the next comment draft TR for OpenMP 6.0:



* Improved specification for Fortran allocatable mapping. The current
  spec comes with limitations when a user wants to map derived types
  that is 'declare target' and that has many allocatable components.



* Better control over mapping a "descriptor" versus the target of the
  descriptor for pointers and allocatables.



* Allow more directives to be combined into a single combined
  directive.



* Support ad-hoc memory partitioning with OpenMP memory allocators.



* Support depend and affinity clauses on 'taskloop' directive.