To: J3                                                     J3/24-156
From: Jeff Larkin
Subject: OpenACC Liaison Report
Date: 2024-June-27

Reference: 22-194

OpenACC Liaison Report
======================
Information about OpenACC, including the current standard document,
training materials, and upcoming events can be found at
http://www.openacc.org/.

The current OpenACC standard is version 3.3. Version 3.3 of the OpenACC
specification was adopted November 2022.  The specification can be found
at https://www.openacc.org/specification. The OpenACC technical committee
is tentatively planning to release a specification update in November
2024.

News ---- The 3.3 specification release focused primarily on standardizing
extensions that were introduced by the NVIDIA compilers and used in
multiple important applications. Several important new features are
detailed below.

* Multiple dimensions of gang parallelism. "Gangs" represent the coarsest
  of OpenACC's 3 levels of parallelism. Several important applications
  reported a challenge in generating sufficient parallelism for modern
  GPUs due to rich loop nests spanning multiple subroutines in the
  callstack. These subroutines prevent easily collpasing the loop nest to
  express more coarse-grained parallelism, so this feature enabled the
  programmer to identify loops in multiple subroutines for coarse-grained
  parallelization. The specification now allows for up to 3 degrees of
  gang parallelism.
* A "force" modifier for the "collapse" clause. Traditionally collapsing
  was only allowed on tightly-nested loops, but this modifier enables the
  programmer to express a desire to collapse nested loops even if
  intervening code appears between those loops. Because this requires
  redundant execution of intervening code, it was deemed best that this
  collapsing be opt-in with a clause modifier.
* Fortran interfaces were added to several memory management routines that
  did not have Fortran bindings previously. These routines were originally
  designed for C-style pointers to device memory and had no corresponding
  functionality in Fortran. Because multiple Fortran OpenACC applications
  were creating their own bindings to these routines it was decided that
  demand was sufficient to standardize on an interface.

Hewlett Packard Enterprise has has continued to update its support for
OpenACC in the crayftn compiler. OpenACC is supported on GPUs from NVIDIA
and AMD as well as multicore targets.

NVIDIA has released HPC SDK 24.5 with support for OpenACC in C, C++, and
Fortran. The 23.11 release introduced a new memory mode that is targeted
at machines with unified memory, either in  hardware (NVIDIA's Grace
Hopper Superchip) or software (Linux HMM). The NVIDIA compilers support
OpenACC for NVIDIA GPUs and multicore CPUs.

GCC supports OpenACC on AMD GPUs and NVIDIA GPUs.

Progress continues on upstreaming support for OpenACC into the LLVM
Compiler Infrastructure Project. This will include support in Fortran and
C/C++.

Applications ------------ There are over 250 applications in production or
development using OpenACC, including:
 - Gaussian 16
 - ANSYS Fluent
 - VASP 6
 - MPAS-A
 - COSMO
 - GAMERA for GPU
 - Quantum Espresso
 - OVERFLOW

More information on these can be found on the OpenACC web site at
https://www.openacc.org/success-stories.

Hackathons ---------- OpenACC continues to host numerous hackathons each
year. OpenACC is generally the preferred approach for Fortran applications
attending the events. Event dates are posted at
https://www.gpuhackathons.org/events. It is not a requirement to use
OpenACC directives to participate in these hackathons.