[petsc-dev] PetscOptionsGetViewer() problems with multiple usage on the same object but different options databases

[Barry Smith]

  One can turn on a multitude of viewers with ad hoc names at the same time like -ksp_view_solution.

  My eventual hope is to consolidate them with an additional part of the argument, for 
example -ksp_view :::solution then one could have -ksp_view :::solution:cancel and something like -ksp_view :::ALL:cancel to cancel all of them.

 But I'd like a short term solution for the current design bug, so I'll eventually try a minimal change to the code.



> On Jan 22, 2023, at 7:59 PM, Jed Brown <jed at jedbrown.org> wrote:
> 
> Can we have a -xxx_view none or null to unset the view? Could even be a viewer type that does nothing. I don't like the out-of-band -xxx_view_cancel.
> 
> Barry Smith <bsmith at petsc.dev> writes:
> 
>>  ALL the other options are sticky.  Calling setFromOptions() twice does not change the value of rtol etc etc etc that was set the last time and even -ksp_monitor is sticky, why are the viewers so special in that they are the only ones that are not sticky? Is not consistent, doesn't make sense, when -ksp_monitor is sticky. It is just a freak feature of the implementation of PetscOptionsGetViewer() that no one realized was there when they wrote it (probably me :-)).
>> 
>> 
>> 
>> 
>> 
>>> On Jan 22, 2023, at 1:16 PM, Jed Brown <jed at jedbrown.org> wrote:
>>> 
>>> Matthew Knepley <knepley at gmail.com> writes:
>>> 
>>>> On Sun, Jan 22, 2023 at 1:06 PM Barry Smith <bsmith at petsc.dev> wrote:
>>>> 
>>>>> 
>>>>> That makes incrementally adding new options to an existing object
>>>>> difficult, since each call to setFromOptions() screws up previous calls.
>>>>> 
>>>>> We already have
>>>>> 
>>>>> -ksp_monitor_cancel
>>>>> -ksp_converged_reason_view_cancel
>>>>> 
>>>> 
>>>> Let me address it another way. If I understand correctly, what happens
>>>> currently is that at each ViewFromOptions() call, the database is inspected
>>>> and if a *_view is found, then the view is executed. So when you replace
>>>> the database, if this option is not found then nothing happens. This seems
>>>> to satisfy a kind of options locality. If instead we make some options
>>>> sticky, like -ksp_view, so that once given they never expire, this would
>>>> make the code exceedingly hard to reason about, especially if options are
>>>> set deep inside something, before a user is knowingly doing things. This
>>>> does not sound like the right model to me.
>>> 
>>> We could have something like PetscOptionsMerge() if one wants to preserve previously set options while adding some new ones (without tainting the global database). I don't like stickiness in general (a key reason why CMake is impossible to reason about and best practice is to run rm -rf * && cmake .. -DALL_THE_OPTIONS).