Re: [eigen] 3.1.4 & 3.0.7 on Friday?

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

On Thu, Aug 1, 2013 at 10:24 PM, Jitse Niesen <jitse@xxxxxxxxxxxxxxxxx> wrote:
> On Thu, 1 Aug 2013, Rhys Ulerich wrote:
>
>>>> Running 'EIGEN_SEED=1375367469
>>>> ./build/test/product_trmm_21' causes a repeatable failure for all
>>>> compilers tested.
>>
>>
>>> I cannot reproduce. Does it fails for this seed only?
>>
>>
>> It failed on several seeds for me.  Using 'while
>> ./build/test/product_trmm_21 ; do echo WORKED ; done' will let your
>> CPU crank on it until either you find one that breaks or you get bored
>> of waiting.
>
>
> I can reproduce it, with the same seed, both for the new 3.1.4 and the
> current dev branch.
>
> I haven't looked at it in detail yet, but the test compares two ways of
> computing A - B, where A is a given matrix and B is a product involving a
> triangular matrix. For some reason, there is a lot of cancellation going on,
> even though these are big random matrices. The norm of A and B are both
> about 300, but the norm of A - B is 0.019 and norm of the difference of
> computing A - B with both methods is 6e-5 (here, norm is the Kronecker norm,
> the sqrt of sum of squares). The test compares 6e-5 with 0.019 and fails,
> but arguably it should compare it with 300 and succeed.

Yes I came with the same conclusion with a different seed.
Surprisingly the cancellation come from the last difference, not the
product. So this seems to fix the issue:

https://bitbucket.org/eigen/eigen/commits/9de6bcf8bd66/


gael

>
> So my guess is that the failure is nothing serious, but it's a tad
> preliminary at the moment, and I need to stop now.
>
>
> Jitse
>
>