This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Improves __ieee754_exp() performance by greater than 5x on sparc/x86.


On 10/20/2017 8:38 AM, Wilco Dijkstra wrote:
Patrick wrote:

         +      fe_val = __fegetround();

Failing to use __fegetround(),__fesetround() causes over 40 math test
accuracy failures for other rounding modes in exp, cexp, cpow, cosh,
ccos, ccosh, csin, and csinh.  If the glibc/Linux community were to
declare that the only rounding mode that was fully supported was
FE_TONEAREST, we could simplify/speed up a lot of code. :-)
"Failure" is too unspecific. What is the difference in ULP? If it is from
say 2 to 20 then yes that's an issue. If it is from 2 to 3 in a non-nearest
rounding mode then that's perfectly acceptable.

Anyway, never use __fegetround/__fesetround. Use get_rounding_mode
which is inlined on all targets and libc_fesetround which is inlined on most
targets.

Wilco


The ULP diffs were mostly in the 2 to 3 range for the complex functions.
There no gross errors.

I appreciate the pointer to get_rounding_mode and libc_fesetround.
I was not previously aware of those functions.
I will give those a try today.

- patrick


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]