[PATCH] math: Add more inputs to atan2 accuracy tests [BZ #28765]

[Sunil Pandey]

Hi Paul,

Thank you so much for running massive accuracy testing on libmvec
function and reporting the issues to us.

I added proposed inputs to v2 and regenerated ulps as requested. Let
me know if it looks ok to you.

Thank you so much for your help,
Sunil


On Wed, Jan 12, 2022 at 8:38 PM Paul Zimmermann
<Paul.Zimmermann@inria.fr> wrote:
>
>        Dear Sunil,
>
> first, with your atan2 patch fixing BZ #28765, I cannot get errors larger than
> 4 ulps with the SSE4.2 code. The largest error I get so far is 2.39063 ulps
> with 0x1.a11891ec004d4p-348,0x1.814830510be26p-348.
>
> > --- a/sysdeps/x86_64/fpu/libm-test-ulps
> > +++ b/sysdeps/x86_64/fpu/libm-test-ulps
> > @@ -214,10 +214,10 @@ double: 1
> >  float: 2
> >
> >  Function: "atan2_vlen4_avx2":
> > -double: 1
> > +double: 2
>
> I wonder whether this should be increased to 3, since I get an error of 3.46942
> ulps for the following inputs and AVX2 (and that inputs should be
> added to math/auto-libm-test-in): 0x1.a83f842ef3f73p-633,0x1.a799d8a6677ep-633
>
> >  Function: "atan2_vlen8":
> > -double: 1
> > +double: 2
> >  float: 2
> >
> >  Function: "atan2_vlen8_avx2":
>
> Best regards,
> Paul
>