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

[Boris Lau <lau@xxxxxxxxxxxxxxxxxxxxxxxxxx>]

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