[PATCH 0/3] Fix powf inaccuracies up to ~169 ULP reported by Paul Zimmermann

[Jeff Johnston]

Hi Fabian,

Thanks for the patches.  They have been merged.

-- Jeff J.

On Wed, Sep 11, 2024 at 10:14 AM Fabian Schriever <
fabian.schriever@gtd-gmbh.de> wrote:

> This patch series fixes the powf accuracy errors reported by Paul
> Zimmermann on the 6th of August.
>
> All three patches are taken from FreeBSD and authored by Bruce Evans, not
> me. The original commits are referenced in the individual patches. The
> patches also contain the original messages.
>
> Having run our test setup the error is reduced from 169 ULP to just shy of
> 1 ULP (0.9998...). Note however that, while quite exhaustive, this is not a
> full range test, and there still may be unobserved errors above 1 ULP.
>
> Fabian Schriever (3):
>   powf: Fix the hi+lo decomposition for 2/(3ln2) (FreeBSD)
>   powf: Fixed 2 bugs in the computation /* t_h=ax+bp[k] High */.
>     (FreeBSD)
>   powf: Fixed another precision bug in powf() (FreeBSD)
>
>  newlib/libm/math/ef_pow.c | 9 +++++----
>  1 file changed, 5 insertions(+), 4 deletions(-)
>
> --
> 2.33.0.windows.1
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://sourceware.org/pipermail/newlib/attachments/20240919/1d03a595/attachment.htm>