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

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

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
>
>
>