Re: [eigen] 3.2-rc1 released!

[ Thread Index | Date Index | More lists.tuxfamily.org/eigen Archives ]


ICE fixed:

https://bitbucket.org/eigen/eigen/commits/68159011f70f/
Changeset:   68159011f70f
Date:        2013-07-19 11:46:54
Summary:     Fix ICE with ICC 11

https://bitbucket.org/eigen/eigen/commits/9393897636eb/
Changeset:   9393897636eb
Branch:      3.2
Date:        2013-07-19 11:46:54
Summary:     Fix ICE with ICC 11
(transplanted from 68159011f70f0997c7e309acb08f1c0aaec6c27d)

Gael

On Fri, Jul 19, 2013 at 11:24 AM, Gael Guennebaud
<gael.guennebaud@xxxxxxxxx> wrote:
> OK, regarding the ICE in unsupported/test/sparse_extra.cpp
>
> - It was introduced by changeset: 39429a449b68
>
> - For the record, with 3.2-beta1 this unit-test produced this strange
> message from ICC 11:
>
> An internal threshold was exceeded for routine
> _Z14sparse_productIN5Eigen19DynamicSparseMatrixIfLi1EiEEEvv and
> optimization level may be reduced. See
> http://software.intel.com/en-us/articles/internal-threshold-was-exceeded
> for more information and advice.
>
> - I managed to workaround this ICE by moving the
> operator*(SparseBase,SparseSelfAdjointView) outside
> SparseSelfAdjointView. However this change makes the official
> test/sparse_product.cpp fails with the same ICE !! So that's even
> worse.
>
> gael
>
> On Fri, Jul 19, 2013 at 9:20 AM, Gael Guennebaud
> <gael.guennebaud@xxxxxxxxx> wrote:
>> On Fri, Jul 19, 2013 at 4:50 AM, Rhys Ulerich <rhys.ulerich@xxxxxxxxx> wrote:
>>>> Eigen 3.2-rc1 has been released today (July 17, 2013).
>>>
>>> I ran 3.2-rc1 Nightly tests run atop an x86_64 GNU/Linux system with
>>> -DEIGEN_TEST_OPENMP=ON for gcc 4.5, gcc 4.6, gcc 4.7, intel 11.1,
>>> intel 12.1, and intel 13.1.  All results are under machine
>>> setun.ices.utexas.edu on the dashboard at
>>> http://manao.inria.fr/CDash/index.php?project=Eigen&date=2013-07-17&display=project.
>>>
>>> On gcc 4.5 I see several test timings like
>>>     223/643 Test #223: product_trmm_11 ..................   Passed   25.42 sec
>>> which look fishy to me.  I am using -DCMAKE_BUILD_TYPE=RelWithDebInfo.
>>>  Most tests are less than 2 seconds and many less than a 10th of a
>>> second.  product_trmm_* and lu_7 seem to be among the outliers.
>>
>> Yes observed that too, investigated a bit but I did not found anything
>> special. These TRMM tests are heavy because there are numerous
>> configuration that have to be tested and the matrices have to be
>> larger than the blocking sizes to make sure everything is well tested.
>> An idea would be to reduce these block-sizes for this test such that
>> we can reduce matrix sizes as well.
>>
>>> On intel 11.1 there are some quite a few failed tests including a few
>>> internal icpc error lovelies.  Is it possible to keep the 3.2 release
>>> from regressing any tests that had worked on 3.1?
>>
>> Yes, we are aware of that ICE too. It's concerning an unsupported
>> unit-test, so not a blocker for me. We could try to bissect it to see
>> if there might be an easy fix though.
>>
>> gael
>>
>>> Hope that helps,
>>> Rhys
>>>
>>>



Mail converted by MHonArc 2.6.19+ http://listengine.tuxfamily.org/