dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Jun 23, 2008 at 12:55:09PM +0100, Garth N. Wells wrote:
> On Jun 23 2008, Anders Logg wrote:
>
>> On Mon, Jun 23, 2008 at 01:18:29PM +0200, Martin Sandve Alnæs wrote:
>>> 2008/6/23 Garth N. Wells <gnw20@xxxxxxxxx>:
>>> >
>>> >
>>> > Anders Logg wrote:
>>> >> On Sat, Jun 21, 2008 at 05:29:14AM -0400, Hatef Monajemi wrote:
>>> >>> Hi Isn't this code supposed to work? It gives me "invalid lvalue 
>>> in >>> assignment" error while compiling. How should I assign the 
>>> elements >>> of a Matrix???
>>> >>>
>>> >>> //-------------------------------
>>> >>> int main()
>>> >>> {
>>> >>>
>>> >>>
>>> >>> for(i=0; i<nens; i++)
>>> >>> .
>>> >>> .
>>> >>> .
>>> >>> real* values = new real[nstat];
>>> >>> u1.interpolate(values);
>>> >>>
>>> >>> Matrix S(nstat,nens);
>>> >>>
>>> >>> for(int j=0;j<nstat;j++)
>>> >>>   {
>>> >>>     S(j,i)=values[j];
>>> >>>   }
>>> >>>
>>> >>> }
>>> >>
>>> >> Sorry, this doesn't work yet.
>>> >>
>>> >> We can easily fix this in the GenericMatrix interface by adding an
>>> >> auxiliary class MatrixEntry (which just holds (i, j) and a reference
>>> >> to the matrix) and have operator() return that, then overload
>>> >> operator= in MatrixEntry to assign to the matrix.
>>> >>
>>> >> We had this in DOLFIN 2-3 years back but the linear algebra has gone
>>> >> through several reimplementations since then...
>>> >>
>>> >> Until this has been implemented, you can use the setitem() function.
>>> >>
>>> >
>>> > I don't know if we want this because the operation is ill-suited to 
>>> > some linear algebra back-ends and adding it encourages its use. If  
>>> > element-wise access is needed, the uBlas matrix class should be 
>>> used.
>>> >
>>> > Garth
>>>
>>> I agree with Garth here. There's also no chance at all that this will
>>> be efficient. It's just not within the scope of the dolfin linear
>>> algebra classes.
>>
>> It's very natural to be able to set values in a matrix. I'm not sure
>> we must always make it impossible to do operations which are not
>> efficient. We can easily discourage it by issuing a warning when
>> matrices are accessed item by item.
>
> It's not impossible to set values - a user just needs to choose the  
> appropriate class, which in this case is the uBlas matrix.
>
> Garth

It's possible also for PETSc, but much less efficient.

-- 
Anders

Attachment: signature.asc
Description: Digital signature