ffc team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Mon, Apr 25, 2011 at 03:14:45PM -0700, Johan Hake wrote:
> On Monday April 25 2011 15:04:43 Anders Logg wrote:
> > On Mon, Apr 25, 2011 at 10:56:25PM +0100, Garth N. Wells wrote:
> > > On 25/04/11 22:48, Anders Logg wrote:
> > > > On Mon, Apr 25, 2011 at 10:41:58PM +0100, Garth N. Wells wrote:
> > > >> On 25/04/11 22:33, Anders Logg wrote:
> > > >>> On Mon, Apr 25, 2011 at 10:26:18PM +0100, Garth N. Wells wrote:
> > > >>>> On 25/04/11 22:08, Anders Logg wrote:
> > > >>>>> On Mon, Apr 25, 2011 at 07:40:21PM -0000, Garth Wells wrote:
> > > >>>>>> On 25/04/11 20:00, Johan Hake wrote:
> > > >>>>>>> On Monday April 25 2011 11:26:36 Garth Wells wrote:
> > > >>>>>>>> On 25/04/11 18:51, Anders Logg wrote:
> > > >>>>>>>>> On Mon, Apr 25, 2011 at 05:11:41PM -0000, Garth Wells wrote:
> > > >>>>>>>>>> On 25/04/11 17:53, Johan Hake wrote:
> > > >>>>>>>>>>> On Monday April 25 2011 08:59:18 Garth Wells wrote:
> > > >>>>>>>>>>>> On 25/04/11 16:47, Johan Hake wrote:
> > > >>>>>>>>>>>>> Commenting out the cache is really not a fix. The problem
> > > >>>>>>>>>>>>> is within dolfin. Isn't there another way to deal with
> > > >>>>>>>>>>>>> this?
> > > >>>>>>>>>>>>
> > > >>>>>>>>>>>> It is a fix if the cache isn't needed.
> > > >>>>>>>>>>>
> > > >>>>>>>>>>> Sure.
> > > >>>>>>>>>>>
> > > >>>>>>>>>>>>> First: How much penalty are there with a disabled memory
> > > >>>>>>>>>>>>> cache. Maybe the problem isn't that bad?
> > > >>>>>>>>>>>>
> > > >>>>>>>>>>>> I don't get the point of this cache. The way it is now, a
> > > >>>>>>>>>>>> form is only preprocessed if it hasn't already been
> > > >>>>>>>>>>>> preprocessed, which seems ok to me. The old code tried to
> > > >>>>>>>>>>>> avoid some preprocessing, but it was highly dubious and I
> > > >>>>>>>>>>>> doubt that it was effective.
> > > >>>>>>>>>>>
> > > >>>>>>>>>>> I think the preprocessing stage actually do take some time.
> > > >>>>>>>>>>> AFAIK the preproces stage essentially do two things. It
> > > >>>>>>>>>>> creates a canonical version of the Form so two Forms that
> > > >>>>>>>>>>> are the same, but constructed at different times are beeing
> > > >>>>>>>>>>> treated equal wrt form generation. Then are DOLFIN specific
> > > >>>>>>>>>>> guys extracted. I am not sure what takes the most time. We
> > > >>>>>>>>>>> should probably profiel it... But if it is the latter we
> > > >>>>>>>>>>> could consider putting another cache in place which is more
> > > >>>>>>>>>>> robust wrt changing DOLFIN objects.
> > > >>>>>>>>>>
> > > >>>>>>>>>> It should be easy to avoid the overhead of preprocessing by
> > > >>>>>>>>>> keeping the object in scope. If the object changes, the only
> > > >>>>>>>>>> robust way to make sure that the form is the same as one in
> > > >>>>>>>>>> the cache is to compare all the data. This requires
> > > >>>>>>>>>> preprocessing the form, which then defeats the purpose of a
> > > >>>>>>>>>> cache. It may be possible to add a lightweight preprocess to
> > > >>>>>>>>>> UFL, but I don't think that it's worth the effort or extra
> > > >>>>>>>>>> complication.
> > > >>>>>>>
> > > >>>>>>> I think a light weight version might be the way to go. This is
> > > >>>>>>> then stored in memory cache. If we are able to strip such a form
> > > >>>>>>> for all DOLFIN specific things we would also prevent huge memory
> > > >>>>>>> leaks with mesh beeing kept.
> > > >>>>>>>
> > > >>>>>>> Then we always grab DOLFIN specific data from the passed form
> > > >>>>>>> instead of grabbing from the cache. Not sure how easy this will
> > > >>>>>>> be to implement, but I think we need to explore it, as the
> > > >>>>>>> DOLFIN specific part of the form really has nothing to do with
> > > >>>>>>> the generated Form.
> > > >>>>>>>
> > > >>>>>>> Martin:
> > > >>>>>>> Why is it important to have the _count in the repr of the form? I
> > > >>>>>>> guess that is used in ufl algorithms? Would it be possible to
> > > >>>>>>> include a second repr function, which did not include the count?
> > > >>>>>>> This would then be used when the signature is checked for. We
> > > >>>>>>> could then use that repr to generate a form which is stored in
> > > >>>>>>> the memory cache. This would then be tripped for any DOLFIN
> > > >>>>>>> specific objects. This should work as the _count attribute has
> > > >>>>>>> nothing to do with what code gets generated, but it is essential
> > > >>>>>>> for internal UFL algorithms, right?
> > > >>>>>>>
> > > >>>>>>>>> I'm not very happy with this change.
> > > >>>>>>>>
> > > >>>>>>>> The bright side is that slow and correct is a better starting
> > > >>>>>>>> point than fast but wrong ;).
> > > >>>>>>>>
> > > >>>>>>>> An easy fix is to attach the preprocessed form to a Form object.
> > > >>>>>>>> This would work robustly if we can make forms immutable once
> > > >>>>>>>> they've been compiled. Is it possible to make a Python object
> > > >>>>>>>> immutable?
> > > >>>>>>>
> > > >>>>>>> We can probably overload all setattribtue methods which prohibits
> > > >>>>>>> a user to write to these but it might not be possible to
> > > >>>>>>> prohibit a user to change attributes on instances owned by the
> > > >>>>>>> Form. I guess this is similare to the difficulties of preserving
> > > >>>>>>> constness in C++, but I think it is even harder in Python.
> > > >>>>>>
> > > >>>>>> What if we have the FFC jit compiler return the preprocessed form,
> > > >>>>>> and inside dolfin.Form simply do
> > > >>>>>>
> > > >>>>>>     class Form(cpp.Form):
> > > >>>>>>         def __init__(self, form, . . .. )
> > > >>>>>>         ....
> > > >>>>>>
> > > >>>>>>         (...., preprocessed_form) = jit(form, . . . . )
> > > >>>>>>
> > > >>>>>>         form = preprocessed_form
> > > >>>>>>
> > > >>>>>>         .....
> > > >>>>>>
> > > >>>>>> This way, form will have form_data, and the FFC jit function will
> > > >>>>>> know not to call ufl.preprocess.
> > > >>>>>
> > > >>>>> Here's another strange thing. In the JITObject class, we have two
> > > >>>>> functions: __hash__ and signature. As far as I understand, the
> > > >>>>> first is used to located objects (generated code/modules) in the
> > > >>>>> Instant in-memory cache, while the second is used for the on-disk
> > > >>>>> cache.
> > > >>>>>
> > > >>>>> >From some simple tests I did now, it looks like the __hash__
> > > >>>>> >function
> > > >>>>>
> > > >>>>> does not need to any significant speedup. The JIT benchmark runs
> > > >>>>> just as fast if I call signature from within __hash__.
> > > >>>>>
> > > >>>>> Furthermore, the __hash__ function must also be broken since it
> > > >>>>> relies on calling id on the form.
> > > >>>>>
> > > >>>>> Ideally, we should get Instant to handle the caching, both
> > > >>>>> in-memory and on-disk, by providing two functions __hash__ (fast,
> > > >>>>> for in-memory cache) and signature (slow, for on-disk cache).
> > > >>>>>
> > > >>>>> Since __hash__ cannot call id, it must be able to attach a unique
> > > >>>>> string to the form (perhaps based on an internal counter in FFC).
> > > >>>>> My suggestion would be to add this to UFL, something like set_hash
> > > >>>>> and hash (which would return None if set_hash has not been called).
> > > >>>>> If Martin does not like that, we should be able to handle it on the
> > > >>>>> DOLFIN side.
> > > >>>>>
> > > >>>>> So in conclusion: no in-memory cache in FFC (handled by Instant)
> > > >>>>> and FFC attaches a hash to incoming forms so that Instant may
> > > >>>>> recognize them later.
> > > >>>>
> > > >>>> The code that I disabled was caching preprocessed forms, so I don't
> > > >>>> see how this can be handled by Instant.
> > > >>>
> > > >>> The point would be that one could check that "hash" of the form (some
> > > >>> unique string) instead of computing the signature which involves
> > > >>> preprocessing the form.
> > > >>
> > > >> How would the hash be computed? To check if the mesh has changed, my
> > > >> limited understanding is that the entire object would have to be
> > > >> serialised, and then a hash computed. How expensive is that?
> > > >>
> > > >> The issue that I ran into was not related to signatures. It was
> > > >> related to the non-UFL data that is attached to arguments.
> > > >
> > > > The hash would be unique to each form. It could just be a counter
> > > > value and the counter would be increased inside Instant for each
> > > > object it gets as input.
> > >
> > > But how does Instant know if a form is new? I also don't see why Instant
> > > should need to know if the mesh associated with a form has changed, but
> > > is for the rest the same. Wouldn't Instant need to be DOLFIN-aware?
> >
> > The hash() function would play the same role as the id() function
> > before with the difference that we can't get the same id for a new
> > form as for an old form that's gone out of scope.
> >
> > Instant should not need to know anything it just does this:
> >
> >    check if object has a set_hash() function
> >    if so, calls hash() to get the hash value
> >      checks the cache for that hash value
> >    if not, assign unique value by calling set_hash on the object
> >
> > We would need to make sure from the DOLFIN side that when we change a
> > Form, we also change the hash value (for example by setting it to
> > None) which would trigger the Instant disk cache.
>
> Sounds complicated...

I think it sounds very easy. Everything we need is there: Instant
already has memory and disk cache. We just need to provide the proper
input.

> Now the preprocessed form is stored in the original form. This will never
> change. Whenever a form does not go out of scope the preprocessed form will
> live.
>
> Also Martin made it impossible to change a form without returning a new
> instance. This prevents any changing of the original form while keeping a
> preprocesses form attached to it.
>
> If a form has a preprocessed form that will be used for code generation. The
> preprocessed form will be used in instants memory cache. The preprocessed form
> has nothing to do the any DOLFIN objects that comes with the original form,
> such as mesh, expressions and such.
>
> Anything I have missed?

What about the __hash__ function in jitobject.py? It still calls
id(). Isn't that a problem?

--
Anders