[petsc-users] about MATSOR

[Sun, Hui]

Thank you Barry. It is a step in forming my PC matrix for the schur complement. I think I will try something else. 

Best,
Hui


________________________________________
From: Barry Smith [bsmith at mcs.anl.gov]
Sent: Wednesday, February 25, 2015 1:12 PM
To: Sun, Hui
Cc: petsc-users at mcs.anl.gov
Subject: Re: [petsc-users] about MATSOR

> On Feb 25, 2015, at 3:02 PM, Sun, Hui <hus003 at ucsd.edu> wrote:
>
> Thank you, Barry. In fact, my matrix B is sparse. Is it possible to drop the terms less than a certain threshold from C, so that C can be sparse?

  You can do that but it becomes much more complicated and expensive, especially assembling back the global sparse matrix.   Why do you want to do this? Are you doing some algebraic multigrid method with smoothed agglomeration or something? If so, this is big projection and it would likely be better to build off some existing tool rather than writing something from scratch.

>
> Another question, there are pc_sor_its and pc_sor_lits in MatSOR. What is local its?

   In parallel it does its "global iterations" with communication between each iteration and inside each global iteration it does lits local iterations (without communication between them). (Some people call this hybrid parallel SOR.

   Sequentially the number of SOR iterations it does is its*lits since there is never any communication between processes since there is only one.

Barry

>
> Best,
> Hui
>
> ________________________________________
> From: Barry Smith [bsmith at mcs.anl.gov]
> Sent: Wednesday, February 25, 2015 11:48 AM
> To: Sun, Hui
> Cc: petsc-users at mcs.anl.gov
> Subject: Re: [petsc-users] about MATSOR
>
>   Let me try to understand what you wish to do. You have a sequential matrix A and a  matrix B and you wish to compute the (dense) matrix C where each column of C is obtained by running three iterations of SOR (using the matrix A) with the corresponding column of B as the right hand side for the SOR; using an initial guess of zero to start the SOR?
>
>   If this is the case then create a seq dense matrix C of the same size as B (it will automatically be initialized with zeros).  Say B has n rows
>
>   VecCreateSeq(PETSC_COMM_SELF,n,&b);
>   VecCreateSeqWithArray(PETSC_COMM_SELF,1,n,NULL,&x);
>   PetscScalar *xx,*carray;
>   MatCreateSeqDense(PETSC_COMM_SELF,n,n,NULL,&C);
>   MatDenseGetArray(C,&carray);
>   loop over columns
>      ierr = MatGetColumnVector(B,b,col);CHKERRQ(ierr);
>      ierr = VecPlaceArray(x,carray + n*col);    /* this makes the vec x point to the correct column of entries in the matrix C.
>> err = MatSOR(A, b, 1, (SOR_FORWARD_SWEEP|SOR_ZERO_INITIAL_GUESS), 0, 3, 0, x);CHKERRQ(ierr);
>
>
>   Barry
>
> Note you could also do this in parallel but it is kind of bogus be the SOR is run independently on each process so I don't recommend it as useful.
>>
>
>> On Feb 25, 2015, at 12:14 PM, Sun, Hui <hus003 at ucsd.edu> wrote:
>>
>> I want to do 3 steps of gauss seidel from a Mat A to another Mat B. Is there a way to do this?
>>
>> I mean, what I can think of is to get the column vectors of B by
>> ierr = MatGetColumnVector(B,v,col);CHKERRQ(ierr);
>>
>>
>>
>> and apply Gauss seidel from A to v:
>> ierr = MatSOR(A, v, 1, (SOR_FORWARD_SWEEP|SOR_ZERO_INITIAL_GUESS), 0, 3, 0, x);CHKERRQ(ierr);
>>
>>
>>
>> But then I have to form another matrix C whose columns are composed by v, and I'm not sure how to do that.
>>
>>
>>
>> Best,
>>
>> Hui
>>