==========================
* New features:
+ Add support for the MPI-4.0 standard.
- Use large count MPI-4 routines.
- Add persistent collective communication.
- Add partitioned point-to-point communication.
- Add new communicator constructors.
- Add the `Session` class and its methods.
+ Add support for the MPI-4.1 standard.
- Add non-destructive completion test for multiple requests.
- Add value-index datatype constructor.
- Add communicator/session buffer attach/detach/flush.
- Support for removal of error classes/codes/strings.
- Support for querying hardware resource information.
+ Add preliminary support for the upcoming MPI-5.0 standard.
- User-level failure mitigation (ULFM).
+ `mpi4py.util.pool`: New drop-in replacement for `multiprocessing.pool`.
+ `mpi4py.util.sync`: New synchronization utilities.
+ Add runtime check for mismatch between `mpiexec` and MPI library.
+ Support `scikit-build-core`_ as an alternative build backend.
.. _scikit-build-core: https://scikit-build.readthedocs.io/
+ Support `meson-python`_ as an alternative build backend.
.. _meson-python: https://meson-python.readthedocs.io/
* Enhancements:
+ `mpi4py.futures`: Support for parallel tasks.
+ `mpi4py.futures`: Report exception tracebacks in workers.
+ `mpi4py.util.pkl5`: Add support for collective communication.
+ Add methods `Datatype.fromcode()`, `Datatype.tocode()` and
attributes `Datatype.typestr`, `Datatype.typechar` to simplify
NumPy interoperability for simple cases.
+ Add methods `Comm.Create_errhandler()`, `Win.Create_errhandler()`,
and `File.Create_errhandler()` to create custom error handlers.
+ Add support for pickle serialization of instances of MPI types.
All instances of `Datatype`, `Info`, and `Status` can be
serialized. Instances of `Op` can be serialized only if created
through `mpi4py` by calling `Op.Create()`. Instances of other MPI
types can be serialized only if they reference predefined handles.
+ Add `handle` attribute and `fromhandle()` class method to MPI
classes to ease interoperability with external code. The handle
value is an unsigned integer guaranteed to fit on the platform's
``uintptr_t`` C type.
+ Add lowercase `free()` method to MPI classes to ease MPI object
deallocation and cleanup. This method eventually attempts to call
`Free()`, but only if the object's MPI handle is not a null or
predefined handle, and such call is allowed within the World Model
init/finalize.
* Backward-incompatible changes:
+ Python 2 is no longer supported, Python 3.6+ is required, but
typing stubs are supported for Python 3.8+.
+ The `Intracomm.Create_group()` method is no longer defined in the
base `Comm` class.
+ `Group.Compare()` and `Comm.Compare()` are no longer class methods
but instance methods. Existing codes using the former class
methods are expected to continue working.
+ `Group.Translate_ranks()` is no longer a class method but an
instance method. Existing codes using the former class method are
expected to continue working.
+ The `LB` and `UB` datatypes are no longer available, use
`Datatype.Create_resized()` instead.
+ The `HOST` predefined attribute key is no longer available.
+ The `MPI.memory` class has been renamed to `MPI.buffer`. The old
name is still available as an alias to the new name.
+ The `mpi4py.dl` module is no longer available.
+ The `mpi4py.get_config` function returns an empty dictionary.
* Miscellaneous:
+ The project is now licensed under the BSD-3-Clause license. This
change is fairly inconsequential for users and distributors. It
simply adds an additional clause against using contributor names
for promotional purposes without their consent.
+ Add a new guidelines section to documentation laying out new fair
play rules. These rules ask companies and outside developers to
refrain from reusing the ``mpi4py`` name in unaffiliated projects,
publishing binary mpi4py wheels on the main Python Package Index
(PyPI), and distributing modified versions with incompatible or
extended API changes. The primary motivation of these rules is to
avoid fragmentation and end-user confusion.