dolfin team mailing list archive

[Martin Sandve Alnæs <martinal@xxxxxxxxx>]

I fixed the test so the near test now passes on 32bit.
The test failures there now are something different.

About 'eps should be scaled': sure. That would be great.
With the < to <= change I basically just did the smallest possible
change in the implementation of near to make near(x,x) == True.

But using abs(x-x0) < eps may be a better of the simple solutions.
It should certainly make near(x,x) == True.

Using relative error would be better but more expensive, and if somebody
tries that they better read your linked article well and do a good job...

Martin

On 8 July 2011 13:10, Garth N. Wells <gnw20@xxxxxxxxx> wrote:
> Here's an extensive discussion on comparing floats:
>
> http://www.cygnus-software.com/papers/comparingfloats/comparingfloats.htm
>
> Garth
>
> On 08/07/11 11:44, Garth N. Wells wrote:
>>
>>
>> On 08/07/11 11:33, Martin Sandve Alnæs wrote:
>>> I already explained this, but let me make it clearer:
>>>
>>> If
>>>   x >> eps,
>>> such that
>>>   x + eps == x,
>>> then near(x, x) implemented as
>>>   x < x + eps
>>> will be equivalent to
>>>   x < x
>>> which is false.
>>>
>>
>> eps should be scaled.
>>
>>> Clearly, near(x,x) should always be true in a correct implementation.
>>> Using x <= x + eps fixes this. This is not "relying on equality",
>>> that would be using ==.
>>>
>>> If you have a failing test case, show it to me and I will see what I can do.
>>>
>>
>> See buildbot.
>>
>> Garth
>>
>>> Martin
>>>
>>>
>>> On 8 July 2011 12:21, Garth N. Wells <gnw20@xxxxxxxxx> wrote:
>>>> I think that this is a 'bug' in the test. Relying on equality (i.e. <=
>>>> versus <) for non-integers will lead to problems somewhere (like the
>>>> 32-bit buildbot failing).
>>>>
>>>> Garth
>>>>
>>>>
>>>> On 08/07/11 09:23, noreply@xxxxxxxxxxxxx wrote:
>>>>> ------------------------------------------------------------
>>>>> revno: 6054
>>>>> committer: Martin Alnes <martinal@xxxxxxxxx>
>>>>> branch nick: work
>>>>> timestamp: Fri 2011-07-08 10:20:54 +0200
>>>>> message:
>>>>>   Re-introduce bugfix in near after recent mistaken(?) re-introduction of the bug, or something... Also disable debug print in test.
>>>>> modified:
>>>>>   dolfin/math/basic.cpp
>>>>>   test/unit/math/python/basic.py
>>>>>
>>>>>
>>>>> --
>>>>> lp:dolfin
>>>>> https://code.launchpad.net/~dolfin-core/dolfin/main
>>>>>
>>>>> Your team DOLFIN Core Team is subscribed to branch lp:dolfin.
>>>>> To unsubscribe from this branch go to https://code.launchpad.net/~dolfin-core/dolfin/main/+edit-subscription
>>>>
>>>> _______________________________________________
>>>> Mailing list: https://launchpad.net/~dolfin
>>>> Post to     : dolfin@xxxxxxxxxxxxxxxxxxx
>>>> Unsubscribe : https://launchpad.net/~dolfin
>>>> More help   : https://help.launchpad.net/ListHelp
>>>>
>>
>>
>> _______________________________________________
>> Mailing list: https://launchpad.net/~dolfin
>> Post to     : dolfin@xxxxxxxxxxxxxxxxxxx
>> Unsubscribe : https://launchpad.net/~dolfin
>> More help   : https://help.launchpad.net/ListHelp
>
> _______________________________________________
> Mailing list: https://launchpad.net/~dolfin
> Post to     : dolfin@xxxxxxxxxxxxxxxxxxx
> Unsubscribe : https://launchpad.net/~dolfin
> More help   : https://help.launchpad.net/ListHelp
>