[petsc-dev] VecNest

[Matthew Knepley]

On Tue, Jul 12, 2022 at 8:25 AM Pierre Jolivet <pierre at joliv.et> wrote:

>
>
> On 12 Jul 2022, at 2:50 PM, Matthew Knepley <knepley at gmail.com> wrote:
>
> On Tue, Jul 12, 2022 at 7:41 AM Matthew Knepley <knepley at gmail.com> wrote:
>
>> On Tue, Jul 12, 2022 at 7:39 AM Pierre Jolivet <pierre at joliv.et> wrote:
>>
>>>
>>>
>>> On 12 Jul 2022, at 2:32 PM, Matthew Knepley <knepley at gmail.com> wrote:
>>>
>>> On Mon, Jul 11, 2022 at 1:17 PM Pierre Jolivet <pierre at joliv.et> wrote:
>>>
>>>> Hello,
>>>> Could anyone help me understand what is going on in the following
>>>> example, please?
>>>> I have a VecNest.
>>>> I either: a) initialize all values to 0.0, then set a specific part of
>>>> the vector to nonzero or b) initialize a part of the vector to 0.0 and set
>>>> the other part to nonzero.
>>>> I don’t see why a) and b) produce different results.
>>>>
>>>> $ ./ex1111 -pc_type fieldsplit -ksp_monitor_true_residual
>>>> -ksp_converged_reason -fieldsplit_pc_type jacobi -ksp_pc_side right
>>>> -ksp_view_final_residual -nest_subvec true
>>>>   0 KSP unpreconditioned resid norm 8.375635517980e-01 true resid norm
>>>> 8.375635517980e-01 ||r(i)||/||b|| 1.000000000000e+00
>>>>   1 KSP unpreconditioned resid norm 4.748816884247e-01 true resid norm
>>>> 4.748816884247e-01 ||r(i)||/||b|| 5.669798875623e-01
>>>>   2 KSP unpreconditioned resid norm 4.713006778679e-01 true resid norm
>>>> 4.713006778679e-01 ||r(i)||/||b|| 5.627043784990e-01
>>>>   3 KSP unpreconditioned resid norm 7.092979927129e-02 true resid norm
>>>> 7.092979927129e-02 ||r(i)||/||b|| 8.468587144106e-02
>>>>   4 KSP unpreconditioned resid norm 1.457836310255e-02 true resid norm
>>>> 1.457836310255e-02 ||r(i)||/||b|| 1.740567992870e-02
>>>>   5 KSP unpreconditioned resid norm 1.625040500524e-14 true resid norm
>>>> 1.633468028779e-14 ||r(i)||/||b|| 1.950261595401e-14
>>>> Linear solve converged due to CONVERGED_RTOL iterations 5
>>>> KSP final norm of residual 1.63347e-14
>>>> $ ./ex1111 -pc_type fieldsplit -ksp_monitor_true_residual
>>>> -ksp_converged_reason -fieldsplit_pc_type jacobi -ksp_pc_side right
>>>> -ksp_view_final_residual -nest_subvec false
>>>>   0 KSP unpreconditioned resid norm 0.000000000000e+00 true resid norm
>>>> 8.375635517980e-01 ||r(i)||/||b||            inf
>>>> Linear solve converged due to CONVERGED_ATOL iterations 0
>>>> KSP final norm of residual 0.837564
>>>>
>>>
>>> I find if I assemble the vector, I get the same answers. Will try to
>>> figure out what assembly is doing.
>>>
>>>
>>> It’s probably reseting all these values
>>> https://petsc.org/main/src/vec/vec/interface/rvector.c.html#line511,
>>> which I believe are being used in VecNorm() inside VecNormalize().
>>> I guess any call to VecNestSubVec() should invalidate those as well,
>>> otherwise we get wrong cached norms.
>>> I will give this a go.
>>>
>>
>> I believe the bug is the following:
>>
>>   You change values in a subvector, which does StateIncrease on the
>> subvector, but not StateIncrease on the nest vector, so it has its cached
>> norm
>>
>> I am not sure what to do about this, since how can the parent know you
>> pulled out the subvector? Will think.
>>
>
> Okay, the error is more subtle. VecSet() automatically computes some
> norms. VecNestGetSubVec() is broken with respect to caching.
> I think the right fix now is VecNestGetSubVec/s() should increase the
> state no matter what. If we want to get more precise, we need
> to have Restore() for the subvecs, which checks the state of the subvector
> when it is restored.
>
> Do you want to change it?
>
>
> https://gitlab.com/petsc/petsc/-/merge_requests/5420 gets the job done,
> and is hopefully not too far from what you had in mind.
> If anyone is not confortable with the incurred cost (e.g., of recomputing
> some norms), please let me know.
>

This is great. Thanks for finding it.

   Matt


> Thanks,
> Pierre
>
>   Thanks,
>
>      Matt
>
>
>>    Matt
>>
>>
>>> Thanks,
>>> Pierre
>>>
>>>   Thanks,
>>>
>>>     Matt
>>>
>>>
>>>> Thanks,
>>>> Pierre
>>>>
>>>>
>>>
>>> --
>>> What most experimenters take for granted before they begin their
>>> experiments is infinitely more interesting than any results to which their
>>> experiments lead.
>>> -- Norbert Wiener
>>>
>>> https://www.cse.buffalo.edu/~knepley/
>>> <http://www.cse.buffalo.edu/~knepley/>
>>>
>>>
>>>
>>
>> --
>> What most experimenters take for granted before they begin their
>> experiments is infinitely more interesting than any results to which their
>> experiments lead.
>> -- Norbert Wiener
>>
>> https://www.cse.buffalo.edu/~knepley/
>> <http://www.cse.buffalo.edu/~knepley/>
>>
>
>
> --
> What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> -- Norbert Wiener
>
> https://www.cse.buffalo.edu/~knepley/
> <http://www.cse.buffalo.edu/~knepley/>
>
>
>

-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener

https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20220712/f4c242d4/attachment-0001.html>