Re: [eigen] Incorrect result when inverting 4x4 Matrix with gcc7.3 and -ffast-math

[Patrik Huber <patrikhuber@xxxxxxxxx>]

Hi Michael,

You could test whether this has anything to do with SSE/AVX. So you could play around with a few -march=... flags - for example "core2" for SSE3 only, or "native" for "all your CPU has got". You can also selectively add e.g.  -msse4.2 or -mavx.

I don't think it has anything to do with your "problem" with -ffast-math but I incidentially found some problems with >=gcc-6 as well this month (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84280).

I would also encourage you to report this on the gcc bugtracker to let them decide whether this is a bug or not. You can send an email to register (the email address should be on https://gcc.gnu.org/bugzilla/). If your email gets rejected for spam reasons (happened to me when I registered...) let me know and I can (indirectly) help.

Best regards,

Patrik

On 28 February 2018 at 11:50, Dan Čermák <dan.cermak@xxxxxxxxxxxxxxxxxxx> wrote:

Hi Michael,

I am not an Eigen developer, so I cannot give you a definitive answer,
but afaik -ffast-math enables the compiler to aggressively reorder
floating point operation to increase performance. That comes at the cost
of loosing IEEE compliance. I have myself seen numerical code break
spectacularly with -ffast-math because it relies the floating point
operations being precise.

So I my guess is that you are seeing a new optimization that gcc enables
backfiring and if anything is at fault, then maybe gcc.


Cheers,

Dan

Michael Mößner <michael.moessner@xxxxxx> writes:

> Hello
>
> I get incorrect results, when I invert a 4x4 matrix with gcc7.3 and the
> compiler option -ffast-math. Unfortunatly, I don't get access to the bug
> tracker, so I post it here.
>
> Example code: (compile with "g++ test.cpp -O3 -ffast-math")
>
> #include <Eigen/Dense>
> #include <iostream>
> using namespace std;
>
> int main(int argc, char **argv)
> {
>      Eigen::Matrix<double,4,4> test;
>      test(0,0) = 1.; test(0,1) = 0.; test(0,2) = 0.; test(0,3) = 0.;
>      test(1,0) = 0.; test(1,1) = 1.; test(1,2) = 0.; test(1,3) = 0.;
>      test(2,0) = 0.; test(2,1) = 0.; test(2,2) = 1.; test(2,3) = 0.;
>      test(3,0) = 0.; test(3,1) = 0.; test(3,2) = 0.; test(3,3) = 1.;
>
>      cerr << test.inverse() << endl;
>      return 1;
> }
>
> ------------------------------
>
> This results in
>
> 1  0  0  0
> -0 -1 -0 -0
>   0  0  1  0
> -0 -0 -0 -1
>
> ------------------------------
>
> The expected result is
>
> 1 0 0 0
> 0 1 0 0
> 0 0 1 0
> 0 0 0 1
>
> ---------------------------------
>
> I also tested this code with gcc4.8.5, gcc5.2 and clang6.0.0 where the
> problem does NOT occur.
>
> I also tested with gcc7.2 where the problem occurs.
>
> The problem does NOT occur when ommitting the -ffast-math or when
> compiling with -O0
>
> My Eigen version:
>
> 3.3.4
>
> My system data:
>
> SUSE Linux Enterprise Desktop 12 SP2
>
> Intel(R) Xeon(R) CPU E3-1276 v3 @ 3.60GHz
>
> ---------------------------
>
> I also tested the following systems where the bug also occurs:
>
> ----------------------------
>
> SUSE Linux Enterprise Desktop 11 SP2
>
> Intel(R) Xeon(R) CPU E3-1276 v3 @ 3.60GHz
>
> ------------------------------
>
> SUSE Linux Enterprise Desktop 12 SP2
>
> Intel(R) Xeon(R) CPU E3-1230 @ 3.20GHz
>
> ------------------------------
>
> Best regards
>
> Michael

--

Dr. Patrik Huber
Centre for Vision, Speech and Signal Processing
University of Surrey
Guildford, Surrey GU2 7XH
United Kingdom

Web: www.patrikhuber.ch
Mobile: +44 (0)7482 633 934