dolfin team mailing list archive

[kent-and@xxxxxxxxx]

>
>
> kent-and@xxxxxxxxx wrote:
>>>
>>> Anders Logg wrote:
>>>> On Wed, Aug 20, 2008 at 04:17:11PM +0200, kent-and@xxxxxxxxx wrote:
>>>>>> kent-and@xxxxxxxxx wrote:
>>>>>>>> Can Assembler::assemble_system be easily generalised to an Array
>>>>>>>> of
>>>>>>>> DirichletBC,
>>>>>>>>
>>>>>>>>    Array<DirichletBC*> bcs
>>>>>>>>
>>>>>>>> so that vector equations can be handled?
>>>>>>>>
>>>>>>>> Garth
>>>>>>>>
>>>>>>> Good point, I'll have a look.
>>>>>>>
>>>>>> OK, I'm taking a look at the moment with a view to generalising
>>>>>> Assembler::assemble_system for nonlinear problems (Newton solve).
>>>>>>
>>>>>> Garth
>>>>> OK, in the present implementation, the matrix does not have 1 on the
>>>>> diagonal
>>>>> associated with bcs. It will be 1 times the number of elements (or
>>>>> faces)
>>>>> associated
>>>>> with it.
>>>>>
>>>>> This weighting can be stored but it is not presently done.
>>>>>
>>>>> Kent
>>>> I amazed you even managed to implement that code. It's huge. We should
>>>> try to break it up so we can reuse as much as possible between that
>>>> and the standard assemble() function.
>>>>
>>> I was impressed :).
>>>
>>> Now that what's needed is clear, it can be broken up. It looks to me
>>> like a number of the existing functions can be extended slightly and
>>> then re-used. The actual code for applying the bcs is not very long.
>>>
>>> Garth
>>>
>>
>> It is a mess,  way to large and should be broken up.
>> It is also a bit "difficult" to break up. For instance, the
>> coefficents may or may not need to be computed depending
>> on whether it has been computed earlier ie. after computing coefficients
>> for a macro element it should not be needed when computing the cell
>> integral.
>> The same applies to dof_maps.
>
> Is there so much overhead involved in this? We're probably recomputing
> these things in the other assembly functions. Perhaps there is
> noticeable overhead, and that's why assemble_system is faster?

The other functions recompute. And there are some recomputations
in assemble_systems also. I guess avoiding these recomputations
and fewer matrix insertions is the reason for speed up. I don' know
which matter most.

>
> Hence it is hard to group everything associated
>> with eg a cell integral together in a function. Or maybe it can be done
>> with a set
>> of bools that contains the state of the various structures.
>>
>
> I'm testing now whether an Array of boundary condition can be applied as
> simply as I think. If that works, assemble_system should be fully
> functional and we can try breaking it up.

Great!

Kent