Re: [Bug-gsl] [PATCH] Disable use of FMA

[Rhys Ulerich]

Hi Tuomo,

Can you comment on the problem Andreas is experiencing with some
ode-initval2 test cases?

>>> The algorithms in cblas (at least) become instable when the compiler is
>>> allowed to contract floating-point expressions.  This manifests in the
>>> failure of the ode-initval2 test.

>> When you say "instable" do you mean that they fail at some particular
>> tolerance level?  Could you give an example?  The problem you describe
>> sounds like overly-sensitive unit tests rather than a reason to
>> entirely avoid FMAs.

> This is the failure I get:
>
> gsl: msbdf.c:1336: ERROR: msbdf_apply too large order change
> -- order change -2
>
> This comes from test_extreme_problems, with the "ringmod" problem.
> Enabling DEBUG in msbdf.c shows this difference between
> -ffp-contract=off and -ffp-contact=fast:
>
> --- dstep: 9.81743e-17 3.08687e-48 9.07047e-10 -9.07047e-10 -9.07047e-10 
> 9.07047e-10 -3.62750e-13 -2.20616e-27 -6.93678e-59 -1.81409e-16 -1.81409e-16 
> 1.81409e-16 1.81409e-16 1.57079e-14 4.93899e-46
> +-- dstep: 9.81743e-17 -8.68549e-36 9.07047e-10 -9.07047e-10 -9.07047e-10 
> 9.07047e-10 -3.62750e-13 -2.20616e-27 5.60182e-46 -1.81409e-16 -1.81409e-16 
> 1.81409e-16 1.81409e-16 1.57079e-14 -1.96797e-42

I don't know that region of GSL well enough to understand the
discrepancy between 3.08687e-48 and -8.68549e-36; -6.93678e-59  and
5.60182e-46; or 4.93899e-46 and -1.96797e-42.  All of them seem like
floating point noise near zero to me...

- Rhys