Maybe it makes sense to propose adding a C binding for a built-in datatype that supports C++. While the Forum clearly wants to kill C++ _bindings_, they insist that the C bindings are sufficient for C++, meaning that they still intend to support C++ as a language. There is no contradiction in adding support for C++ datatypes like std::complex while at the same time killing the C++ bindings. I think that they are separate issues and that there is a strong case to be made for adding an MPI datatype for C++ complex that works via the C bindings. Jeff On Tue, Jun 26, 2012 at 12:02 AM, Jed Brown <[email protected]> wrote:

I'm all for getting rid of the C++ bindings and see no compelling reason for anyone to use them.

Predefined types are inconsequential to MPI-1 functionality, but they are special for one-sided, so availability of types is highly relevant to applications. This is a non-cosmetic change imposed by #281 that doesn't seem to have been discussed.

On Jun 25, 2012 8:51 PM, "Fab Tillier" <[email protected]> wrote:

...

Ah, thanks, sorry for being dense (I'm not familiar with the C++ bindings as we don't ship them for MSMPI).

Ticket 281 removes the entire C++ bindings from the standard, MPI::COMPLEX included. �You'd need to convince your MPI library vendor to keep supplying you with an MPI-2.2 compliant version of the C++ bindings (which is still allowed by ticket #281).

Hopefully someone with more experience with the C++ bindings can chime in, but I'd expect the answer to your original question is yes, ticket #281 intentionally removes all C++ support from MPI-3.

-Fab

Jed Brown wrote on Mon, 25 Jun 2012 at 21:31:55

...

MPI::COMPLEX is not the same as MPI_C_COMPLEX (and the latter is not available on systems without C99).

...

On Jun 25, 2012 8:21 PM, "Fab Tillier" <[email protected]> wrote: Hi Jed,

Maybe I'm not following you, but what predefined MPI datatype supports std::complex today? �My understanding is that there is no support for std::complex in MPI today, and that ticket 281 does not change this at all.

Thanks, -Fab

Jed Brown wrote on Mon, 25 Jun 2012 at 20:42:53

...

I see that std::complex will not be accessible in a predefined type if #281 passes in its current form. Do you intend for there to be a supported/recommended way for a C++ caller to get a complex predefined (so that one-sided can be used) type? Note that C99 is not a subset of any C++ and C99 complex cannot be used portably from C++ (even with compiler suites that happen to support both).

On Jun 25, 2012 7:07 PM, "Fab Tillier" <[email protected] <mailto:[email protected]> > wrote: Hi Jed,

Jed Brown wrote on Mon, 25 Jun 2012 at 16:43:54

...

Is it intended that #281 would make std::complex inaccessible? Should there be a MPI_CXX_COMPLEX, MPI_CXX_DOUBLE_COMPLEX, MPI_CXX_LONG_DOUBLE_COMPLEX, and (possibly) MPI_CXX_BOOL?

I don't think ticket #281 does anything with respect to std::complex accessibility. �The MPI standard defines the MPI_C_COMPLEX and family as mapping to the C99 _Complex type (see Table 3.2, Predefined MPI datatypes corresponding to C datatypes). �The standard is clear that there are no corresponding C++ bindings.

...

Note that C99 complex is not a substitute because Microsoft does not implement it and it has different semantics regarding numerical stability. (Well, C99 semantics are specified and useful, std::complex stability is not specified by any C++ standard and in practice, is not implemented in a stable way.) This is not such a big deal for use with collectives because user-defined MPI_Ops can be used, but only predefined types can be used with one-sided operations, so presence of complex datatypes (or lack thereof) is relevant to applications.

The fact that Microsoft does not support C99 is lame. �You can probably get away using std::complex as a stand-in for C99 _Complex (but then you're on your own), or move to a compiler that does (I believe the Intel compiler supports C99 on Windows).

-Fab

_______________________________________________ mpi-forum mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum

-- Jeff Hammond Argonne Leadership Computing Facility University of Chicago Computation Institute [email protected] / (630) 252-5381 http://www.linkedin.com/in/jeffhammond https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond