Hi,
So, I've spent some time investigating. Until now I haven't looked into
the "new" FloatVar code contributed by Vincent and you guys, and it
looks like it inherits some ideas, maybe even files from our initial
implementation for Gecode 2. So, I feel some blame for using Boost's
Interval Arithmetic library without investigating enough first: if you
look at the documentation homepage, you'll see a big red warning :)
http://www.boost.org/doc/libs/1_61_0/libs/numeric/interval/doc/interval.htm
which says that the library probably won't work with SSE2 (or at least
not out of the box). But SSE2 is the only option for x64.
On the other hand, I'm not aware of any good alternative to this Boost
library. I think the library is actually quite good (even though it
would use some work on supporting SSE2 better).
So I thought I'd play with the library's policies to see if I could make
it work with x64. I looked into float.hh which defines what exact
interval datatype is used. It looks like changing from
'rounded_arith_opp' to 'rounded_arith_std' fixes the issues that we're
observing. If you're interested the description of those types is
available here:
http://www.boost.org/doc/libs/1_61_0/libs/numeric/interval/doc/rounding.htm
I've run the Float tests with this change and they all seem to pass.
Compared to the _opp version the total runtime running the Float
testsuite increased by 50%. This is still fine for me; I haven't
compared this with the x86/x87 version. Maybe Gecode could use some
#ifdef to choose between _opp and _std based on whether the
compiler/architecture setup is known to be safe for _opp or not.
Cheers,
Filip
Hi Filip,
Thanks for all the info and checking. I'll have a look a little later.
Best
Christian
--
Christian Schulte, www.gecode.org/~schulte
Professor of Computer Science, KTH, cschu...@kth.se
Expert Researcher, SICS, cschu...@sics.se
-----Original Message-----
From: users-boun...@gecode.org [mailto:users-boun...@gecode.org] On Behalf
Of Filip Konvicka
Sent: Monday, April 4, 2016 09:15
To: users@gecode.org
Subject: Re: [gecode-users] Freeze with MSVC 2015 optimized build
Hi,
You said that you have uploaded packages recently; I assume that it was
Gecode-4.4.0-x64-2015.msi. I compiled and linked using this with MSVC
2015 Update 2 and it also freezes on the testcase. Replacing this with the
MSVC 2013 package (Gecode-4.4.0-x64-2013.msi) works fine, even using the
same toolset.
Cheers,
Filip
Hi,
I have updated to MSVC 2015 update 2, compiled the current Gecode
trunk for x64, and my testcase from the beginning of this thread still
freezes (the one from Sep 3rd 2015 - I have attached it to this mail
for convenience).
I also tried running Float::Arithmetic::Div::Sol::A and that test
still fails. For some reason 'make check' only runs a small subset of
tests, so I ran this manually like this:
$ test/test -test Float::Arithmetic::Div::Sol::A
Float::Arithmetic::Div::Sol::A -
Options: -seed 1 -test Float::Arithmetic::Div::Sol::A
Another try failed with a different seed number:
$ test/test -test Float::Arithmetic::Div Float::Arithmetic::Div::A
+++++ Float::Arithmetic::Div::B +++++ Float::Arithmetic::Div::C +++++
Float::Arithmetic::Div::Sol::A -
Options: -seed 1975575111 -test Float::Arithmetic::Div::Sol::A
Thanks,
Filip
Christian,
Thanks. I was referring to this September & October thread - there
were issues with Gecode freezing on some FloatVar problems (my
testcase and your "Float::Arithmetic::Div::C") with MSVC 2015, and
you said then that it might be resolved by not using SSE2 (but that
Guido would investigate more). We also identified a number of float
tests that were failing on GCC 5 (not with GCC 4.9).
I can try re-running our tests here with the current trunk to see if
something changed - I just thought that I would ask directly whether
you did something specific to address these issues. The failures
seemed a bit non-deterministic...
Thanks,
Filip
MSVC2015 works for Gecode.4.4.0 since its initial release, I in fact
uploaded packages to the Gecode website last week. Since last week
(Mar 30) the update 2 to MSVC 2015 finally also compiles the trunk
and the upcoming Gecode 5.0.0.
Don't know about GCC, though.
Cheers
Christian
--
Christian Schulte, www.gecode.org/~schulte Professor of Computer
Science, KTH, cschu...@kth.se Expert Researcher, SICS,
cschu...@sics.se
-----Original Message-----
From: users-boun...@gecode.org
[mailto:users-boun...@gecode.org] On Behalf Of Filip Konvicka
Sent: Friday, April 1, 2016 14:41
To: users@gecode.org
Subject: Re: [gecode-users] Freeze with MSVC 2015 optimized build
Hi Christian,
Have you had any chance to look into supporting gcc 5 and MSVC 2015?
Thanks,
Filip
Hi Filip,
Thanks for the information. This really looks like a gcc problem.
Guido, can you reproduce this behavior?
Cheers
Christian
--
Christian Schulte, www.gecode.org/~schulte Professor of Computer
Science, KTH, cschu...@kth.se Expert Researcher, SICS,
cschu...@sics.se
-----Original Message-----
From: users-boun...@gecode.org
[mailto:users-boun...@gecode.org] On Behalf Of Filip Konvicka
Sent: Friday, October 2, 2015 09:58 AM
To: users@gecode.org
Subject: Re: [gecode-users] Freeze with MSVC 2015 optimized build
Hi,
We were looking a bit more into 4.4.0 on Linux where we've also hit
some issues.
All tests worked fine with gcc 4.9.3. However, with gcc 5.2.0 many
tests failed - see the attachements for details:
gecode-4.4.0-gcc-5.2.0-failed_tests.list
... list of failed tests (except two memory fails below)
gecode-4.4.0-gcc-5.2.0-failed_tests.log
... log from 'test' for failed tests
Two tests ran out of memory (allocated several GBs of memory before
getting killed by system):
Set::Branch::Sparse::3
Set::Channel::Sorted
gcc details:
target=x86_64-linux-gnu
GCC_VER=5.2.0
GMP_VER=6.0.0
MPFR_VER=3.1.3
MPC_VER=1.0.3
BINUTILS_VER=2.25
It could also be a gcc bug of course, hard to tell.
Thanks,
Filip
_______________________________________________
Gecode users mailing list
users@gecode.org
https://www.gecode.org/mailman/listinfo/gecode-users
_______________________________________________
Gecode users mailing list
users@gecode.org
https://www.gecode.org/mailman/listinfo/gecode-users
_______________________________________________
Gecode users mailing list
users@gecode.org
https://www.gecode.org/mailman/listinfo/gecode-users
