On 30 November 2011 00:56, Marie E. Rognes<meg@xxxxxxxxx> wrote:
On 29. nov. 2011, at 21:11, Kristian Ølgaard<k.b.oelgaard@xxxxxxxxx> wrote:
On 29 November 2011 14:00, Anders Logg<logg@xxxxxxxxx> wrote:
On Tue, Nov 29, 2011 at 01:49:15PM +0100, Johannes Ring wrote:
On Tue, Nov 29, 2011 at 1:37 PM, Anders Logg<logg@xxxxxxxxx> wrote:
On Tue, Nov 29, 2011 at 12:15:31PM +0100, Johannes Ring wrote:
On Tue, Nov 29, 2011 at 7:15 AM, Anders Logg<logg@xxxxxxxxx> wrote:
The buildbot is broken after yesterday's release. The release script
was partly broken so I ended up redoing the release a couple of timees
before everything worked.
The strange thing now is that the tests work fine locally in my
repository. I have committed and bzr seems to think I have nothing to
push. Can someone else check? In particular, do you have 1.0-rc1+ in
the reference header files? I do.
I have 1.0-rc1 in the files in test/regression/references/r_auto, but
in test/regression/references/r_quadrature* I have version 1.0-beta2+.
Strange!
Look at this:
$ bzr pull lp:ffc
No revisions to pull.
$ bzr push lp:ffc
No new revisions to push.
$ bzr diff lp:ffc
$ grep "generated by" test/regression/references/r_auto/Poisson.h
// and was automatically generated by FFC version 1.0-rc1+.
Sorry, I meant I have 1.0-rc1+ in test/regression/references/r_auto:
$ grep "generated by" test/regression/references/r_auto/Poisson.h
// and was automatically generated by FFC version 1.0-rc1+.
But I have 1.0-beta2+ in test/regression/references/r_quadrature*:
$ grep "generated by" test/regression/references/r_quadrature/Poisson.h
// and was automatically generated by FFC version 1.0-beta2+.
$ grep "generated by" test/regression/references/r_quadrature_O/Poisson.h
// and was automatically generated by FFC version 1.0-beta2+.
Do you?
Same here. ok, now I see what the problem is. The recent work by Marie
of adding regression testing for various compiler options means the
release script needs to be updated.
We're now running
python test.py --fast
which means not all references will be updated.
Should we add another flag in the release script, or change --fast to
include all representations?
Opinions from Marie and Kristian?
I think that --fast should be changed to include all representations.
The current meaning of --fast is to compare the generated header files
to the references, but to skip program validation (g++ compilation,
executing code and comparing output values). Now we have references
for two more representations (not just 'auto') so the natural thing to
do is to compare the output of those representations too. It will be a
bit slower, but still a lot faster than running all tests.
I think fast should be kept as actually fast, iow as is.
But what is --fast used for, how do you use it?
I usually run the full regression test before pushing anyway.
I could use --fast if I add minor changes to the workflow inside FFC
which I either expect to have either no effect or a specific effect on
the output.
In both cases, however, it would be enough to compare the output of a
few forms like [Poisson, VectorLaplaceGradCurl, etc.], i.e., making
--fast --> --really-fast.
Perhaps, it could be useful to enable testing of single forms?
We have now added generate-only which compares the generated header files for all chosen parameter sets, but does not build, run and compare the output of the auxiliary test problems. On second thought, we might want to change that to compile-only if desired.
I think generate-only is OK, we do both FFC compile and GCC compile in
the regression tests, so 'compile-only' is ambiguous.
