[petsc-users] preallocation for FE matrix

[Tabrez Ali]

You are right. I overlooked using integer(8) for [nnzpr].

Works fine with PetscInt.

Thanks


On Oct 4, 2010, at 4:50 PM, Barry Smith wrote:

>
>  Then perhaps your calls to MatSetValues() are wrong? Everything  
> cannot be write and yet still get the wrong preallocation.
>
>   Here is a trick you can do after providing the preallocation info  
> to the matrix but before calling MatSetValues() call  
> MatSetOption(mat,MAT_NEW_NONZERO_LOCATION_ERR) now it will  
> automatically stop when it finds it has "run out of preallocated  
> spaces" during a set values so in the debugger you can see what has  
> happened when the "impossible" actually happens.
>
>   Barry
>
>
> On Oct 4, 2010, at 4:43 PM, Tabrez Ali wrote:
>
>> Barry Smith wrote:
>>>
>>> On Oct 4, 2010, at 4:21 PM, Tabrez Ali wrote:
>>>
>>>
>>>> I am trying to assemble a FE matrix (adding one element at a time  
>>>> from the local stiffness matrix) and cant seem to get the  
>>>> preallocation right. I am correctly calculating the number of non  
>>>> zeros per row and storing the value in the array [nnzpr].
>>>>
>>>
>>>  (For sequential case) you cannot be calculating the number of  
>>> nonzeros per row correctly and yet not getting the preallocation  
>>> correct.
>>>
>>>  For your simple problem. Print the nonzeros per row you are  
>>> computing then print the sparse matrix you computing. That will  
>>> show which rows are not computed or set properly.
>>>
>>
>> I am calculating the nonzeros per row correctly and have checked  
>> this by explicitly forming the dense matrix.
>>
>> Here is the stiffness matrix (1 corresponds to non zero location)  
>> pattern for the same problem.
>>
>>  1  1  1  1  1  1  1  1  0  0  0  0
>>  1  1  1  1  1  1  1  1  0  0  0  0
>>  1  1  1  1  1  1  1  1  1  1  1  1
>>  1  1  1  1  1  1  1  1  1  1  1  1
>>  1  1  1  1  1  1  1  1  1  1  1  1
>>  1  1  1  1  1  1  1  1  1  1  1  1
>>  1  1  1  1  1  1  1  1  0  0  0  0
>>  1  1  1  1  1  1  1  1  0  0  0  0
>>  0  0  1  1  1  1  0  0  1  1  1  1
>>  0  0  1  1  1  1  0  0  1  1  1  1
>>  0  0  1  1  1  1  0  0  1  1  1  1
>>  0  0  1  1  1  1  0  0  1  1  1  1
>>
>> and here is the [nnzpr] array that I passed to MatCreateSeqBAIJ
>>
>>   8
>>   8
>>  12
>>  12
>>  12
>>  12
>>   8
>>   8
>>   8
>>   8
>>   8
>>   8
>>
>> The sum of values in [nnzpr] adds up to 112.
>>
>> Thanks again
>>>> I dont understand why -mat_view_info shows allocated nonzeros to  
>>>> be 154 when sum(nnzpr) is 112.
>>>>
>>>
>>>   If you mess up the count on a row it allocates a few extra for  
>>> that row (incase you end up needing them) so the number allocated  
>>> can be more then the number actually needed.
>>>
>>>   Barry
>>>
>>>
>>>
>>>> Here is part of the relevant code
>>>>
>>>> ...
>>>> call MatCreateSeqBAIJ(Petsc_Comm_Self, 1, m, n,  
>>>> petsc_null_integer, nnzpr, Mat_A, ierr)
>>>> ...
>>>> ! Assume a bilinear quad (2 dof per node)
>>>> do j1=1,8
>>>>  do j2=1,8
>>>>    call MatSetValues(Mat_A, 1, indx(j1)-1, 1, indx(j2)-1,  
>>>> k(j1,j2), Add_Values, ierr)
>>>>  end do
>>>> end do
>>>> ...
>>>> call MatAssemblyBegin(Mat_A,Mat_Final_Assembly,ierr)
>>>> call MatAssemblyEnd(Mat_A,Mat_Final_Assembly,ierr)
>>>> ...
>>>>
>>>> On running it I get
>>>>
>>>> -bash-3.00$ ./a.out <inp -info -mat_view_info
>>>> ...
>>>> Total non-zero elements estimated i.e., sum(nnzpr)  
>>>> =                  112
>>>> [0] PetscCommDuplicate(): Duplicating a communicator 1140850689  
>>>> -2080374784 max tags = 2147483647
>>>> [0] PetscCommDuplicate():   returning tag 2147483647
>>>> [0] MatAssemblyEnd_SeqBAIJ(): Matrix size: 12 X 12, block size 1;  
>>>> storage space: 42 unneeded, 112 used
>>>> [0] MatAssemblyEnd_SeqBAIJ(): Number of mallocs during  
>>>> MatSetValues is 9
>>>> [0] MatAssemblyEnd_SeqBAIJ(): Most nonzeros blocks in any row is 12
>>>> ...
>>>> Matrix Object:
>>>>  type=seqbaij, rows=12, cols=12
>>>>  total: nonzeros=112, allocated nonzeros=154
>>>>      block size is 1
>>>>
>>>> I dont understand why -mat_view_info shows allocated nonzeros to  
>>>> be 154 when sum(nnzpr) is 112.
>>>>
>>>> Thanks in advance.
>>>>
>>>
>>>
>>
>