[petsc-users] Shell matrix and MatMultAdd.

[Barry Smith]

On Dec 2, 2010, at 3:02 PM, Vijay S. Mahadevan wrote:

>>   I suggest making a shell MatMultAdd_Mine() that does everything then
>> also code a MatMult_Mine() that zeros the output vector and then calls directly MatMultAdd_Mine(). This way you'll
>> have one code to routine but can handle both operations.
> 
> Barry, thanks for the suggestion. I will implement the MatMultAdd in
> this way. Puristically speaking though, this does reverse the
> dependency between the two routines !

   Well one could argue that MatMult() is a special case of MatMultAdd() with a zero vector input.  

  Barry

> 
> Vijay
> 
> On Thu, Dec 2, 2010 at 2:51 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>> 
>> On Dec 2, 2010, at 2:45 PM, Vijay S. Mahadevan wrote:
>> 
>>> Hi all,
>>> 
>>> There is probably some minor inconsistency in my understanding but is
>>> there a fundamental difference between the following two options ?
>>> 
>>> // Option 1
>>> ierr = MatMult(A, solution, temporaryvec) ;CHKERRQ(ierr);
>>> ierr = VecAXPY(rhs, 1.0, temporaryvec) ;CHKERRQ(ierr);
>>> // Option 2
>>> ierr = MatMultAdd(A, solution, rhs, rhs) ;CHKERRQ(ierr);
>>> 
>>> Here A is a shell matrix (serial) that has a routine defined to
>>> perform MATOP_MULT only.
>>> 
>>> I ask because Option 1 gives me the right result while Option 2
>>> segfaults at the MatMultAdd line. My only logical conclusion is that I
>>> need to define a MATOP_MULT_ADD or something similar for the shell for
>>> this to work. Is this understanding correct ?
>> 
>>   Yes, if your code will use a MatMultAdd() then you need to provide that to the shell matrix.
>> 
>>> I implicitly assumed
>>> that petsc recognizes that MATOP_MULT has been defined already and
>>> since MatMultAdd only requires the action of a matrix on a vector to
>>> perform its operation, should this not be computed by petsc
>>> automatically ?
>> 
>>   Sorry it doesn't though of course it could.
>> 
>>> I really do not want to creat an extra vector here
>>> with Option 1 since this occurs at a finer level in my calculation.
>>> But is there any way that you would suggest I do this without extra
>>> allocations ? Any comments or pointers will be much appreciated.
>> 
>>   I suggest making a shell MatMultAdd_Mine() that does everything then
>> also code a MatMult_Mine() that zeros the output vector and then calls directly MatMultAdd_Mine(). This way you'll
>> have one code to routine but can handle both operations.
>> 
>>   Barry
>> 
>> 
>> 
>>> 
>>> Vijay
>> 
>>