[petsc-users] Is there a bug in VecRestoreSubVector?

[Jed Brown]

If you have a Git clone, you can make a temporary merge to see if it works.

On October 9, 2014 10:06:44 AM MDT, Eric Chamberland <Eric.Chamberland at giref.ulaval.ca> wrote:
>Hi Jed,
>
>I do not have a petsc/dev or petsc/next compiled (I just use petsc 
>3.5.2), but Patrick have one for sure.  He should be able to try the 
>patch sooner than I.
>
>Patrick, would you give it a try please?
>
>Thanks,
>
>Eric
>
>On 10/09/2014 11:41 AM, Jed Brown wrote:
>> Hi, Eric.  Fabian reported this a couple days ago (all, petsc-users
>is
>> preferred so that people can see recent discussion).
>>
>> Thanks for linking to Patrick's report.  I have extended it to
>simplify
>> the code a bit and reduce synchronization, but I don't have a test
>case
>> handy and I'm about to board a plane.  Could you try 'next' (or
>> 'placasse/VecGetSubVector' in petsc.git, which has my follow-up
>commits)
>> to let me know if it works for you.
>>
>>
>https://bitbucket.org/petsc/petsc/issue/75/vecrestoresubvector-silently-ignores-non
>>
>> Eric Chamberland <Eric.Chamberland at giref.ulaval.ca> writes:
>>
>>> Hi,
>>>
>>> I noticed it because I got a strange behavior when computing y=Ax
>with
>>> (a wrongly) non-contiguous stride in parallel.  Normally, the
>>> VecRestoreSubVector should have exited with an error ("Unhandled
>case,
>>> values have been changed and need to be copied back into X) but
>didn't,
>>> resulting in a silently wrong result (ouch).
>>>
>>> I corrected my code to define a contiguous stride, which "bypass"
>the
>>> bug in VecRestoreSubVector and now I hopefully compute the correct
>result.
>>>
>>> You may be interested by:
>>>
>>>
>>>
>https://bitbucket.org/placasse/petsc/commits/4e4323587ad9148c29ba9101e5030629f073700c
>>>
>>> By the way, if I have no local unknowns on a process, shall the
>stride
>>> *always* be considered locally contiguous? (this is my initial error
>on
>>> how to define a stride when no unknowns are locally owned).
>>>
>>> Thanks,
>>>
>>> Eric