[Insight-developers] 0008911: Numerical issues in vnl matrix inversion?

Tom Vercauteren tom.vercauteren at m4x.org
Mon Jul 27 12:46:07 EDT 2009


Hi Bill,

To me, both options seems fine. It would be great to have a feedback
from the bug reporter though.

Then if we choose to change the test and use doubles instead of floats
we can still use the patch for itk::Matrix if some other machine still
shows a failing test.

Tom

On Mon, Jul 27, 2009 at 18:29, Bill Lorensen<bill.lorensen at gmail.com> wrote:
> With your patch, the test passes. However, if I use the original
> itkMatrix.h code but change the matlab-related matrices to double, the
> test also passes. Looking at the matlab matrix, I think computing the
> inverse in double makes sense. Perhaps we should leave itkMatrix alone
> and fix the test to use double?
>
> Bill
>
>
> On Mon, Jul 27, 2009 at 12:00 PM, Bill Lorensen<bill.lorensen at gmail.com> wrote:
>> I'll try it on my gcc4.3 system.
>>
>> On Mon, Jul 27, 2009 at 11:50 AM, Tom
>> Vercauteren<tom.vercauteren at m4x.org> wrote:
>>> Hi Bill,
>>>
>>> I was actually writing an email about that exact matter :)
>>>
>>> A few days ago, I tried to replicate bug 8911
>>> http://www.itk.org/Bug/view.php?id=8911
>>>
>>> I couldn't do it on my machine but since I had the test written, I
>>> decided to commit it. Now I see it failing on a few machines. So
>>> apparently, bug 8911 can be replicated.
>>>
>>> I agree that letting vxl folks fix it would be good. Another option
>>> would be to try a fix on ITK's side (which would be faster). I already
>>> prepared something but cannot currently test whether it works or not.
>>> The tentative patch is attached here and to the bug tracker. Could you
>>> try it on your system?
>>>
>>> Thanks,
>>> Tom
>>>
>>> On Mon, Jul 27, 2009 at 17:40, Bill Lorensen<bill.lorensen at gmail.com> wrote:
>>>> Tom,
>>>>
>>>> I see that you checked in a test that exhibits the failing inversion.
>>>> It fails on Borland, gcc 4.3 and MAC 4.0. Does it pass on your system?
>>>>
>>>> Is there a plan to fix this? Perhaps we should we let the vxl folks fix it?
>>>>
>>>> Bill
>>>>
>>>
>>
>


More information about the Insight-developers mailing list