https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122304
Carson Range <crange76 at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|WAITING |RESOLVED
Resolution|--- |FIXED
--- Comment #10 from Carson Range <crange76 at gmail dot com> ---
(In reply to Andrew Pinski from comment #9)
> One more thing contraction is about the use of `fused multiply add` set of
> instructions which does the multipy and add in infinite precision before
> doing the rounding to the specific precision. This can cause unexpected
> rounding errors especially when code is not expecting this added precision.
>
> and if you look at the expected values and the resulting values they are off
> by a small amount:
> 0.896949828 and: 0.896951199
> but higher than the code's threshold of 1e-7.
Thanks setting `-ffp-contract=off` makes it work! This isn't really my library
but looks pretty important that they have color rounding by my lame-mans gasp
of their project. I'm sure someone smart set that threshold that high for a
reason.
