[petsc-users] Scatter context for MatMult

[Mark F. Adams]

On Aug 17, 2012, at 8:38 PM, John Fettig wrote:

> On Fri, Aug 17, 2012 at 5:03 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>> 
>> On Aug 17, 2012, at 1:48 PM, John Fettig <john.fettig at gmail.com> wrote:
>> 
>>> On Fri, Aug 17, 2012 at 2:33 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
>>>> Why do you want raw access?
>>> 
>>> I'm trying to write my own RAP operator for AMG.
>> 
>>   Why? Our's is not fast enough and you think you can make it better?  You want to support an R that is not the transpose of P?  Regardless why not start with the current source code we have and fork your own routine instead of starting from scratch? Doing a full efficient RAP from scratch is a major project.
> 
> Yes, I'm having problems getting MatPtAP to work efficiently with my
> projection matrices.  It is especially problematic in parallel, and
> I'm not really sure why since it seems to do ok for GAMG.  I'm trying
> to code up a simple RAP currently, and I'll let you know how it goes.
> 

As Barry says you are diving down a rabbit hole here unless you want to really get into it (which is fine).

I think the GAMG projection ops that you were using were unsmoothed which are very sparse and so RAP does not do as much work as with most P (R').  Try the smoothing as I mentioned before and use this as your baseline for the speed of RAP.  This smoothed operator probably looks more like your P than the unsmoothed one.

> John
>