I would like to propose an extension to the collective operators to allow them to be applied on non contiguous types, if and only if these types are composed by a single predefined type. The original matching rules (which operator support a specific predefined type) will continue to apply. As an example, this will allow users to apply a reduction operation on a trapezoidal matrix without having to define their own operator. As a result the reduction operation can benefit from all internal optimizations (such as pipelining) that usually are allowed only on contiguous types. george. On May 13, 2008, at 8:54 AM, Jesper Larsson Traeff wrote:
Here are some items that are perhaps suitable for MPI-2.2
* MPI_Reduce_scatter_block
* (possibly) built-in segmented Scan operations (new binary operators)
* Correction/clarification to semantics of one-sided synchronization when switching between modes (NOTE: the proposal I posted some months back is deficient, as Rolf pointed out to me, and will have to be modified)
I will try to make final proposals before the next meeting
Jesper
On Mon, May 05, 2008 at 03:36:01PM -0500, Rajeev Thakur wrote:
Now that MPI 2.1 is close to completion, it is time to look at issues for MPI 2.2. Please send any items you may have for 2.2 to [email protected] and/or upload them to the wiki (and let us know). I had made a note of the following items from the last meeting that were deferred to 2.2:
* support for long long in maxloc, minloc
* fix meaning of first two sentences in intercommunicator barrier
* Ex 5.8 uses old style MPI_UB
* Many functions say "integer" instead of "nonnegative integer"
* const in C++ bindings
* IN, OUT, and INOUT handle types broken
* MPI_CHAR, MPI::CHAR should be signed char or char?
* Need 64-bit sizes for count in messages
* Sizes of MPI_Aint and MPI_Offset should not be different
* New predefined types MPI_AINT and MPI_OFFSET
* New predefined types corresponding to new types in C99, Fortran 2003, and C++ some year
_______________________________________________ mpi-22 mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-22
mpi-forum mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi-forum