dolfin team mailing list archive

[Martin Sandve Alnæs <martinal@xxxxxxxxx>]

On Fri, Jan 23, 2009 at 2:16 PM, Johan Hake <hake@xxxxxxxxx> wrote:
> On Friday 23 January 2009 14:02:31 kent-and@xxxxxxxxx wrote:
>> > On Fri, Jan 23, 2009 at 1:33 PM,  <kent-and@xxxxxxxxx> wrote:
>> >>> On Fri, Jan 23, 2009 at 12:40:21PM +0100, Johan Hake wrote:
>> >>>> On Friday 23 January 2009 12:15:51 Martin Sandve Alnæs wrote:
>> >>>> > On Fri, Jan 23, 2009 at 12:00 PM, Johan Hake <hake@xxxxxxxxx> wrote:
>> >>>> > > On Friday 23 January 2009 11:54:50 Martin Sandve Alnæs wrote:
>> >>>> > >> On Fri, Jan 23, 2009 at 11:48 AM, Johan Hake <hake@xxxxxxxxx>
>> >>>>
>> >>>> wrote:
>> >>>> > >> > On Friday 23 January 2009 11:27:31 Ola Skavhaug wrote:
>> >>>> > >> >> On Fri, Jan 23, 2009 at 10:32 AM, Johan Hake <hake@xxxxxxxxx>
>> >>>>
>> >>>> wrote:
>> >>>> > >> >> > On Friday 23 January 2009 10:09:12 Martin Sandve Alnæs
>> >>>>
>> >>>> wrote:
>> >>>> > >> >> > > On Fri, Jan 23, 2009 at 9:39 AM, Johannes Ring
>> >>>> > >> >> > > <johannr@xxxxxxxxx>
>> >>>> > >> >> >
>> >>>> > >> >> > wrote:
>> >>>> > >> >> > > > On Thu, January 22, 2009 19:10, Johan Hake wrote:
>> >>>> > >> >> > > >> On Thursday 22 January 2009 17:02:41 A Navaei wrote:
>> >>>> > >> >> > > >>> Johan,
>> >>>> > >> >> > > >>>
>> >>>> > >> >> > > >>> Thanks, instant-clean did the trick! Maybe this should
>> >>>>
>> >>>> be
>> >>>>
>> >>>> > >> >> > > >>> somehow automated.
>> >>>> > >> >> > > >>
>> >>>> > >> >> > > >> Yes, I have thought about it.
>> >>>> > >> >> > > >>
>> >>>> > >> >> > > >> Is it possible to add a call to instant-clean in the
>> >>>>
>> >>>> install
>> >>>>
>> >>>> > >> >> > > >> script of the ffc
>> >>>> > >> >> > > >> and dolfin packages Johannes?
>> >>>> > >> >> > > >
>> >>>> > >> >> > > > Yes, it is possible to add post installation scripts to
>> >>>>
>> >>>> the FFC
>> >>>>
>> >>>> > >> >> > > > and DOLFIN packages (and perhaps Instant) that runs
>> >>>> > >> >> > > > instant-clean. Should I add this?
>> >>>> > >> >> > > >
>> >>>> > >> >> > > > Johannes
>> >>>> > >> >> > >
>> >>>> > >> >> > > Can you add it to SyFi as well please? But we don't want
>> >>>>
>> >>>> this in
>> >>>>
>> >>>> > >> >> > > development versions though, we should at least have an
>> >>>>
>> >>>> option
>> >>>>
>> >>>> > >> >> > > to avoid it.
>> >>>> > >> >> >
>> >>>> > >> >> > I think it will be sufficient to add the automatic cleaning
>> >>>>
>> >>>> in
>> >>>> the
>> >>>>
>> >>>> > >> >> > ubuntu scripts.
>> >>>> > >> >> >
>> >>>> > >> >> > I can add the swig version to the signature generation in
>> >>>>
>> >>>> ffc
>> >>>> and
>> >>>>
>> >>>> > >> >> > dolfin.
>> >>>> > >> >> >
>> >>>> > >> >> > While on the topic, does instant check whether swig is
>> >>>>
>> >>>> installed,
>> >>>>
>> >>>> > >> >> > when a module is built? Also I have on several occasions now
>> >>>> > >> >> > checked for the swig version. Should I put this code in
>> >>>>
>> >>>> instant,
>> >>>>
>> >>>> > >> >> > e.g.
>> >>>> > >> >> >
>> >>>> > >> >> >   check_swig_version("1.3.35")
>> >>>> > >> >> >
>> >>>> > >> >> > It will return false if the current swig version is lesser
>> >>>>
>> >>>> than
>> >>>>
>> >>>> > >> >> > "1.3.35"? I think it would be natural for instant to provide
>> >>>>
>> >>>> such a
>> >>>>
>> >>>> > >> >> > function.
>> >>>> > >> >>
>> >>>> > >> >> There is a related problem that needs to be adressed here. If
>> >>>>
>> >>>> you
>> >>>>
>> >>>> > >> >> build an extension of a module wrapped with version x of swig,
>> >>>>
>> >>>> you
>> >>>>
>> >>>> > >> >> should require the same version of swig in instant. Perhaps
>> >>>>
>> >>>> the
>> >>>>
>> >>>> > >> >> check_swig_version("1.3.35") should only return true if you
>> >>>>
>> >>>> have
>> >>>>
>> >>>> > >> >> exactly that version. Another function, a-la
>> >>>> > >> >>
>> >>>> > >> >> assert_swig_min_version("1.3.35")
>> >>>> > >> >
>> >>>> > >> > This could probably be kept in one function. By default the
>> >>>>
>> >>>> function
>> >>>>
>> >>>> > >> > returns whether a version of swig is equal or greater than the
>> >>>> > >> > prescribed version, and:
>> >>>> > >> >
>> >>>> > >> >   check_swig_version("1.3.35", equal = True)
>> >>>> > >> >
>> >>>> > >> > will return true only if the version is the same as the
>> >>>>
>> >>>> prescribed
>> >>>>
>> >>>> > >> > one?
>> >>>> > >> >
>> >>>> > >> > The assertion functionality will then be handled by the
>> >>>>
>> >>>> function
>> >>>> that
>> >>>>
>> >>>> > >> > use instant, i.e., ffc, sfc, dolfin compile_function a.s.o.
>> >>>> > >>
>> >>>> > >> But FFC isn't compiled with SWIG. Dolfin will have to do that,
>> >>>>
>> >>>> for
>> >>>>
>> >>>> > >> example at import time.
>> >>>> > >
>> >>>> > > No, but it produces code that is compiled with swig. And this code
>> >>>>
>> >>>> should
>> >>>>
>> >>>> > > be compiled with the same swig version that dolfin is compiled
>> >>>>
>> >>>> with.
>> >>>> Then
>> >>>>
>> >>>> > > dolfin need to hand its swig version to ffc when it is using it so
>> >>>>
>> >>>> ffc
>> >>>>
>> >>>> > > can check which version is installed while swigging compiled
>> >>>>
>> >>>> forms.
>> >>>> If
>> >>>>
>> >>>> > > this version is not the same as the handed version from dolfin ->
>> >>>>
>> >>>> raise
>> >>>>
>> >>>> > > error.
>> >>>> >
>> >>>> > In my opinion,
>> >>>> >
>> >>>> >   dolfin gets its swig version
>> >>>> >   dolfin calls instant to check whether it's compatible with the
>> >>>>
>> >>>> current
>> >>>>
>> >>>> > swig
>> >>>> >
>> >>>> > is simpler than
>> >>>> >
>> >>>> >   dolfin gets its swig version
>> >>>> >   dolfin passes its swig version to ffc
>> >>>> >   ffc calls instant to check whether it's compatible with the
>> >>>>
>> >>>> current
>> >>>> swig
>> >>>>
>> >>>> > why pass the info to ffc at all? It would need to be handled in
>> >>>>
>> >>>> dolfin
>> >>>>
>> >>>> > for function compilation anyway. If dolfin checks that the current
>> >>>> > swig is compatible with the version it was compiled with at import
>> >>>> > time, it doesn't have to be checked anywhere else. With your model,
>> >>>> > both DOLFIN, FFC and SFC would need to be modified.
>> >>>>
>> >>>> Sounds reasonable. Dolfin defines it own jit function and the check
>> >>>> could be
>> >>>> added here.
>> >>>>
>> >>>> We will still have the problem for older forms, which are allready in
>> >>>> the
>> >>>> cache and are compiled with an older swig. Putting the swig version in
>> >>>> the
>> >>>> signature would prevent this to happen though.
>> >>>>
>> >>>> Johan
>> >>>
>> >>> I think we should throw as much as possible into the signature, just
>> >>> to make sure. DOLFIN passes some signature prefix to the form compiler
>> >>> (which could be the DOLFIN version and the SWIG version), then the
>> >>> form compiler adds whatever it feels to add to the signature (form
>> >>> signature, form compiler name and form compiler version), then Instant
>> >>> adds whatever it feels like adding.
>> >>>
>> >>> --
>> >>> Anders
>> >>
>> >> Ok, so Instant should add Swig version to the signature. Dolfin and the
>> >
>> > No, as Johan pointed out, Instant can't add to the signature. Instant
>> > will either construct its own signature from the files and options it
>> > gets, or use the provided signature as-is. So the form compilers
>> > should add whatever they need to the signature.
>> >
>> > Martin
>>
>> But the form compilers / Dolfin do not know which SWIG version Instant
>> will use
>> (although this might not be a big issue since it is probably basically the
>> same process
>> that decides on which SWIG to use in both cases).
>
> The form compiler/dolfin can get the present swig version (instant can provide
> a function swig_version()?), which will be used to compile a signature.

Exactly.

>> Anyway, this is really independent of Dolfin and the compilers. What
>> should Instant
>> do eg. if SWIG has been updated since the last time it inlined the exact
>> same code?
>
> I think Instant should just do what it is told to. Any checks need to be done
> in the caller of instant.

Exactly. And we can add the swig version if instant creates the
signature itself.

>> Should it just import the module or regenerate it?
>
> Then ffc/dolfin will provide a different signature and a new module will be
> compiled.

Exactly.

Martin