Re: [eigen] Problem with g++-4.4 -O2 and Eigen3

[Bastien ROUCARIES <roucaries.bastien@xxxxxxxxx>]

On Sun, Jul 17, 2011 at 9:44 PM, Gael Guennebaud
<gael.guennebaud@xxxxxxxxx> wrote:
> see this threads for more details about this bug:
>
> http://listengine.tuxfamily.org/lists.tuxfamily.org/eigen/2011/04/msg00013.html

Too large for gcc bugzilla unfortunatly.

I will try to add a bug

Bastien

> gael
>
> On Sun, Jul 17, 2011 at 11:46 AM, Bastien ROUCARIES
> <roucaries.bastien@xxxxxxxxx> wrote:
>> Do you have a test case for reporting bug ?
>>
>> On Wed, Jun 1, 2011 at 4:42 PM, Gael Guennebaud
>> <gael.guennebaud@xxxxxxxxx> wrote:
>>> I'm afraid there is nothing we can do on our side. It's clearly a
>>> compiler bug. For the record, and for the others, on your initial
>>> example I found that a double value was wrongly modified to -1 in
>>> between function calls, something like:
>>>
>>> foo(1);
>>>
>>> void foo(double alpha)
>>> {
>>>   // here alpha still equal 1)
>>>   bar(alpha);
>>> }
>>>
>>> void bar(double alpha)
>>> {
>>>   // here alpha == -1
>>> }
>>>
>>> and even worse, this behavior was dependent on the order of the .o
>>> files in the linking command line, i.e., the bug only appeared for
>>> some orders.
>>>
>>> Valgrind did not reported any issue.
>>>
>>> All other version of gcc was fine.
>>>
>>> So I'd recommend you to upgrade your compiler. At least you could try
>>> with the latest version of 4.4, i.e., 4.4.6 if that's not already the
>>> one you are using.
>>>
>>> gael
>>>
>>> On Wed, Jun 1, 2011 at 4:29 PM, Boris Lau
>>> <lau@xxxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
>>>> Hi,
>>>> sorry, but no, adding the 1 and/or the noalias doesn't change anything..
>>>>
>>>> Here are some more interesting matrix multiplication results. In short, 3x3
>>>> * 3x9 gives a mess, while 3x9 * 9x3 and 9x3 * 3x9 does not.
>>>>
>>>> Best, Boris
>>>>
>>>> Eigen::Matrix<double,3,9> J;
>>>> J.fill(1.0);
>>>>
>>>> std::cout << J << std::endl;
>>>> 1 1 1 1 1 1 1 1 1
>>>> 1 1 1 1 1 1 1 1 1
>>>> 1 1 1 1 1 1 1 1 1
>>>>
>>>> std::cout << Eigen::Matrix3d::Identity()*J << std::endl;
>>>> -1 -1 -1 -1 -1 -1 -1 -1 -1
>>>> -1 -1 -1 -1 -1 -1 -1 -1 -1
>>>> -1 -1 -1 -1 -1 -1 -1 -1 -1
>>>> wrong!
>>>>
>>>> std::cout << J.transpose()*J << std::endl;
>>>> 3 3 3 3 3 3 3 3 3
>>>> 3 3 3 3 3 3 3 3 3
>>>> 3 3 3 3 3 3 3 3 3
>>>> 3 3 3 3 3 3 3 3 3
>>>> 3 3 3 3 3 3 3 3 3
>>>> 3 3 3 3 3 3 3 3 3
>>>> 3 3 3 3 3 3 3 3 3
>>>> 3 3 3 3 3 3 3 3 3
>>>> 3 3 3 3 3 3 3 3 3
>>>> correct!
>>>>
>>>> std::cout << J*J.transpose() << std::endl;
>>>> 9 9 9
>>>> 9 9 9
>>>> 9 9 9
>>>> correct!
>>>>
>>>> Best, Boris
>>>>
>>>> On 05/29/2011 05:08 PM, Gael Guennebaud wrote:
>>>>>
>>>>> ok, so just like before, the default value of the multiplicative
>>>>> factor alpha gets transformed into a -1.... that's really crasy.
>>>>>
>>>>> Does that change anything if explicitly put a multiplicative factor, e.g.:
>>>>>
>>>>> C.noalias() = 1 * A * B;
>>>>>
>>>>> Regarding the matrix size, this is because for matrices smaller than 8
>>>>> we use another product implementation.
>>>>>
>>>>>
>>>>> gael
>>>>>
>>>>>
>>>>> On Sun, May 29, 2011 at 2:45 PM, Boris Lau
>>>>> <lau@xxxxxxxxxxxxxxxxxxxxxxxxxx>  wrote:
>>>>>>
>>>>>> Hi folks,
>>>>>>
>>>>>> after updating Eigen to the newest version in the repository, I now
>>>>>> compile
>>>>>> my code with Eigen including the "no inlining of the GEBP product kernel"
>>>>>> change by Gael.
>>>>>>
>>>>>> However, still a simple matrix multiplication goes awfully wrong when
>>>>>> using
>>>>>> -O2 in g++-4.4:
>>>>>>
>>>>>> // a code snipped
>>>>>> Eigen::Matrix<double,3,9>  J;
>>>>>> J.fill(1.0);
>>>>>> Eigen::Matrix<double,3,9>  Jprime = Eigen::Matrix3d::Identity() * J;
>>>>>> std::cout<<  "J\n"<<  J<<  std::endl;
>>>>>> std::cout<<  "Jprime\n"<<  Jprime<<  std::endl;
>>>>>>
>>>>>> // output
>>>>>> J
>>>>>> 1 1 1 1 1 1 1 1 1
>>>>>> 1 1 1 1 1 1 1 1 1
>>>>>> 1 1 1 1 1 1 1 1 1
>>>>>> Jprime
>>>>>> -1 -1 -1 -1 -1 -1 -1 -1 -1
>>>>>> -1 -1 -1 -1 -1 -1 -1 -1 -1
>>>>>> -1 -1 -1 -1 -1 -1 -1 -1 -1
>>>>>>
>>>>>> As for Hauke, this only occurs only when used deeper inside a project, in
>>>>>> the main function it works alright. The problem also occurs for a matrix
>>>>>> size of 3x8, and disappears for 3x7 and smaller. Using dynamic sizes does
>>>>>> not help.
>>>>>> I also could not find a working order of .o files during linking.
>>>>>>
>>>>>> Maybe the no onlining fix didn't go far enough?
>>>>>>
>>>>>> Boris
>>>>>>
>>>>>> --
>>>>>> Boris Lau
>>>>>>
>>>>>> Albert-Ludwigs-University Freiburg
>>>>>> Institute of Computer Science, Autonomous Intelligent Systems Group
>>>>>> Georges-Koehler-Allee 079, D-79110 Freiburg, Germany
>>>>>>
>>>>>> Building 079, Room 1005
>>>>>> Phone: +49 761 203-8012 | Mobile: +49 174 9436758
>>>>>> http://www.informatik.uni-freiburg.de/~lau
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>> --
>>>> Boris Lau
>>>>
>>>> Albert-Ludwigs-University Freiburg
>>>> Institute of Computer Science, Autonomous Intelligent Systems Group
>>>> Georges-Koehler-Allee 079, D-79110 Freiburg, Germany
>>>>
>>>> Building 079, Room 1005
>>>> Phone: +49 761 203-8012 | Mobile: +49 174 9436758
>>>> http://www.informatik.uni-freiburg.de/~lau
>>>>
>>>>
>>>>
>>>
>>>
>>>
>>
>>
>>
>
>
>