dolfin team mailing list archive

["Garth N. Wells" <gnw20@xxxxxxxxx>]

On 17/01/11 18:16, Anders Logg wrote:
> On Mon, Jan 17, 2011 at 06:07:19PM +0000, Garth N. Wells wrote:
>>
>>
>> On 17/01/11 18:03, Anders Logg wrote:
>>> On Mon, Jan 17, 2011 at 04:30:23PM +0000, Garth N. Wells wrote:
>>>>
>>>>
>>>> On 17/01/11 16:28, Anders Logg wrote:
>>>>> The following doesn't seem to work in Python any longer:
>>>>>
>>>>> A = Matrix(10, 10)
>>>>>
>>>>> Is matrix initialization broken?
>>>>>
>>>>
>>>> Probably not broken - more like
>>>>
>>>>   A = Matrix(10, 10)
>>>>
>>>> is not supported. Since a Matrix is in general sparse, it doesn't make
>>>> sense to initialise it as above (some backends even insist on the
>>>> sparsity being defined at construction).
>>>>
>>>> Garth
>>>
>>> It would be good to allow simple initialization of a Matrix without
>>> requiring to go through all the hassle of creating a SparsityPattern.
>>>
>>> I generally don't think we should disallow certain operations just
>>> because they are potentially slow.
>>
>> It's not just that they're slow. They cannot be supported by all backends.
> 
> Then we throw an error.
>

Better still, use an appropriate data structure ;). If a user wants a
dense matrix, we should make it easy to use a dense matrix.

The Matrix class in DOLFIN is sparse and the premise should be that it's
distributed. Being consistent in this makes life a lot easier in parallel.

>>> Some operations (like Matrix index
>>> access) will only be performed for toy problems or while testing and
>>> then speed is not very important (since the problem is small anyway).
>>>
>>
>> If it's made available, it will be used. This type of operation is main
>> reason for things breaking in parallel. We have getitem and setitem
>> which are sufficiently unfriendly that hopefully users get the idea that
>> they're for testing only.
> 
> Just because something can be misused, it shouldn't be disallowed.
>

Doesn't sound like a strong argument. It should not be easy to misuse a
library. That's why I can live with get/setitem, but not indexing into a
sparse distributed matrix via A(i, j).

> It would be easy (and more helpful) to add a message the first time
> getitem/setitem is used in a program:
> 
>   Warning: Index access to matrix/vector values is potentially very
>   slow and it breaks in parallel. To disable this warning, set the
>   parameter "warning_index_access" to false.
> 

Yes, I would like to see a set/getitem warning.

Garth

> --
> Anders