Jeff,

The first look at the definitions might indeed be a bit confusing, as it is strange to see overloaded C functions (even though C11 _Generic has been around for a while). It is also not clear when the second variant (using MPI_Count) is available, that is only if C11 _Generic is supported by my compiler. Even worse, a C99 compiler will happily cast an MPI_Count argument passed by value to int without complaint or warning by default. Should we mark in the interface the fact that the MPI_Count overloads are only available in C11? I'm thinking about something similar to cppreference's distinction between C/C++ standard versions, e.g.,


```
int MPI_Send(const void *buf, int count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm)

int MPI_Send(const void *buf, MPI_Count count, MPI_Datatype datatype, int dest, int tag, MPI_Comm comm) [>C11|C++]

int MPI_Send_x(const void *buf, MPI_Count count, MPI_Datatype datatype,
int dest, int tag, MPI_Comm comm)
```

That way it is immediately clear that there is a catch here and users have to be careful. OpenShmem, for example, has that distinction too (although more explicit using paragraphs for different language flavors).

Cheers,
Joseph

On 7/30/19 6:13 PM, Jeff Squyres (jsquyres) via mpi-forum wrote:
_*SHORT VERSION*_

See the attached two PDFs for how "BigCount" will be rendered in the MPI-4.0 PDF (pt2pt chapter and pt2pt/external interfaces in Annex A).

Comments / suggestions are welcome, but you may only make comments / suggestions if you have read all of the "MORE DETAIL" section, below.

_*MORE DETAIL*_

We have split the work of the "BigCount" WG into two parts:

1. Updating all the bindings throughout the entire MPI spec.  This is further sub-divided into two parts:
    1a. Updating the bindings in the chapters of the MPI spec.
    1b. Updating the bindings listing in Annex A.

2. Updating text in various chapters (e.g., Terms&Conventions and Language Bindings) describing BigCount, language consequences, describing how the bindings appear in the rest of the doc, ...etc.

We've incorporated all the feedback from the BigCount Virtual Meeting last week, and I'd like to share samples of how #1a and #1b will look in the MPI 4.0 PDF.

Text for #2 will be read at the September meeting (stay tuned -- that text will be published according to the normal 2-week deadline, etc.).

Attached are 2 documents:

1. The point-to-point chapter rendered in the new style.  Look at the bindings for MPI_SEND and MPI_RECV, for example.

2. Annex A rendered in the new style (only some of the bindings have been updated so far -- such as pt2pt and external interfaces).

If you are confused when looking at these PDFs (e.g., why are there up to 3 C bindings?), and/or if you weren't at the Virtual Meeting / didn't watch the Webex after the fact, the feedback we received from the Virtual Meeting is summarized below.  It is a lot of wordswordswords and is somewhat difficult to follow -- I highly suggest reading the rules below while simultaneously examining the attached PDFs.  The summary below should make much more sense when looking at concrete examples.

-----

The overall intent is:

A. In MPI-4.0, expose "BigCount" functionality through language polymorphism in the C bindings and the mpi_f08 module (we are /_not_/ updating mpif.h or the mpi module).

B. C11 _Generic polymorphism kinda sucks, *and* we're in a transition period where not all C compilers are C11-capable.  Hence, we're exposing up to *3* C bindings per MPI procedure to applications (including explicitly exposing the "_x" variant where relevant).

C. Fortran polymorphism "just works" (skipping a longer explanation here -- ask Rolf/Puri if you're interested in the details), so there will only be up to *2* F08 bindings per MPI procedure.

-----

For each binding in the main text of the document:

1. Show the "plain" MPI_Foo(...) C binding (e.g., as it appeared in MPI-3.1)

2. If the binding has count or displacement arguments:
   2a. Show a second MPI_Foo() C binding with count and displacement arguments as MPI_Count and MPI_Aint, respectively.    2b. If MPI_Foo() DOES NOT HAVE a corresponding MPI_Foo_x() in MPI-3.1, show a third MPI_Foo_x(...) C binding with count/displacement as MPI_Count/MPI_Aint.

3. Show the "plain" MPI_Foo(...) F08 binding (e.g., as it appeared in MPI-3.1)

4. If the binding has count or displacement arguments:
   4a. Show a second MPI_Foo() F08 binding with count and displacement arguments as INTEGER(KIND=MPI_COUNT_KIND) and INTEGER(MPI_ADDRESS_KIND), respectively.

5. Show the "plain" F90 binding

-----

References:

- Slides we presented at the BigCount Virtual Meeting: https://docs.google.com/presentation/d/10D_ws5NYAOI8m6KZKb31s9HxiSSmc_k5C620Cuy3brY/edit

- Webex Recording of the BigCount Virtual Meeting: https://www.mpi-forum.org/virtual_meetings/

- Froozle MPI, a no-op MPI implementation showing a bindings implementation of BigCount: https://github.com/jsquyres/froozle-mpi

If you have read all the way down to this point, comments and feedback are welcome.  :-)

--
Jeff Squyres
jsquy...@cisco.com <mailto:jsquy...@cisco.com>


_______________________________________________
mpi-forum mailing list
mpi-forum@lists.mpi-forum.org
https://lists.mpi-forum.org/mailman/listinfo/mpi-forum

_______________________________________________
mpi-forum mailing list
mpi-forum@lists.mpi-forum.org
https://lists.mpi-forum.org/mailman/listinfo/mpi-forum

Reply via email to