To: J3                                                     J3/20-161
From: Bill Long
Subject: OpenMP Liaison Report
Date: 2020-October-13


OpenMP Report for Meeting 222
-----------------------------

The OpenMP 5.1 specification will be released next month. A comment
draft for OpenMP 5.1 was released this past August.

Since February, there were three virtual "F2F" meetings, in May, July,
and another a week ago. Here is a summary of the major changes from
5.0 to 5.1 (I'm also highlighting the ones that are Fortran-specific):

- Deprecated the master construct. A new construct, masked, was
  added. It has identical semantics to the master construct, but it
  also accepts a "filter" clause that can selects a particular thread
  to execute the structured block of the construct. We also replaced
  references to the "master thread" in the specification and the
  Examples document with "primary thread".

- [Fortran] Clarified how use_device_ptr, use_device_addr, and
  is_device_ptr clauses are supposed to work for Fortran. The
  use_device_ptr and is_device_ptr clauses should permit C_PTR
  variables, and specifying other types of variables is deprecated for
  those clauses.

- [Fortran] Clarified how data-sharing attributes apply to associate
  names.

- [Fortran] Added restrictions for certain uses of coarrays and
  related-features in OpenMP regions. In sum:
      - If a program is running with multiple images, explicit OpenMP
        regions cannot contain image control statements, ERROR STOP
        statement, FAIL IMAGE statement, collective subroutine calls,
        or references to coindexed coarrays.
      - If a coarray is privatized, the private copy cannot be
        coindexed or passed to a procedure with a coarray dummy
        argument.
      - Coarrays cannot be threadprivate.
      - List items in clauses cannot be coindexed coarrays.

- [Fortran] Clarified that OpenMP directives are not permitted inside
  a DO CONCURRENT construct. Have not yet addressed usage of LOCALITY
  clauses in Fortran 2018.

- Added more flexibility to the declare variant and metadirective
  features, allowing dynamic selection of function variants and
  directive variants.

- Added an error directive for issuing compile-time or run-time errors
  or warnings under a specified condition.

- Added a tile and unroll construct. More loop transformation
  constructs are expected in future releases.

- Added support for making indirect calls in an offload region through
  function/procedure pointer.

- Added interoperability support with other non-OpenMP device runtimes
  through a new interop construct. This construct is available for all
  base languages. As an example, one can use this feature to obtain a
  CUDA stream and synchronize work on the stream with OpenMP
  constructs. However, the APIs for obtaining implementation-specific
  properties (like a CUDA stream) from a "foreign" device runtime are
  only available for C/C++.

- Added support for compare-and-swap atomics with the new "compare"
  atomic type for atomic constructs. This atomic type can also be used
  to support min/max atomic operations in C/C++ (min/max atomics are
  already available for Fortran, as they are intrinsic procedures that
  can be used in a standard atomic construct).

- Allowed default(private) and default(firstprivate) clauses in C/C++
  (these were already available in Fortran).

- Added the omp_display_env() routine for displaying environment
  variables. A motivation for this was to allow one to only have 1 MPI
  rank display the environment variables values.

- Added asynchronous memcpy routines for devices
  (omp_target_memcpy_async and omp_target_memcpy_rect_async), for all
  base languages. As with other device memory routines, for Fortran
  these routines use the C interoperability types.

Also, the plan is to release a 5.2 version of the specification
sometime next year. The purpose of that release will not be to add new
features, but to implement improvements to the format of the
specification itself.