[mpich-discuss] Using __float128 with MPI_Accumulate()?

[Pavan Balaji]

On 12/25/2011 09:23 PM, Jeff Hammond wrote:
> The primary reason that the MPI standard does not allow user-defined
> reductions for MPI_Accumulate is not the nature of floating-point
> types, but rather the operations themselves, since user-defined
> accumulate is equivalent to active-messages, which cannot be
> implemented in hardware.  Of course, __float128 is not supported by
> any hardware I know of, but neither is complex double, but the Forum
> found a way to let that one slide.

The concern was not that there currently doesn't exist hardware to do an 
operation X (say complex double), but rather that it's not possible to 
build hardware that does arbitrary user-defined operations.

The easiest way for this is to get float128 defined as a predefined OP 
in MPI.  We added a bunch in MPI-2.2.  It should be fairly 
straightforward to add more even in 3.0 if we can get the text written 
up before this Forum and in time for a formal reading.  If not, 3.1 
should definitely be possible.

  -- Pavan

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji