* faster exp10f code
@ 2020-04-06 11:51 paul zimmermann
2020-04-07 20:36 ` Joseph Myers
0 siblings, 1 reply; 5+ messages in thread
From: paul zimmermann @ 2020-04-06 11:51 UTC (permalink / raw)
To: libc-alpha
Hi,
the 3 patches referenced below implement some faster exp10f code, inspired by
the expf code, that is much faster than the current one (tested on x86_64).
Here are the "make bench" results (I had to add an entry for exp10f):
before new code:
"exp10f": {
"": {
"duration": 3.40525e+09,
"iterations": 6.4368e+07,
"max": 65906.3,
"min": 24.056,
"mean": 52.9029
}
with new code:
"exp10f": {
"": {
"duration": 3.30806e+09,
"iterations": 1.59728e+08,
"max": 39557.8,
"min": 16.104,
"mean": 20.7106
}
I disabled the code in math/w_exp10f_compat.c by adding #if 0 ... #endif
around it, since I don't know how to properly do it.
Paul Zimmermann
[1] https://homepages.loria.fr/PZimmermann/glibc-contrib/0001-added-new-code-for-exp10f-inspired-from-expf-and-reu.patch
[2] https://homepages.loria.fr/PZimmermann/glibc-contrib/0002-e_exp10f.c-cleaned-up-added-error-analysis-and-comme.patch
[3] https://homepages.loria.fr/PZimmermann/glibc-contrib/0003-integrate-new-exp10f-code.patch
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: faster exp10f code
2020-04-06 11:51 faster exp10f code paul zimmermann
@ 2020-04-07 20:36 ` Joseph Myers
2020-04-07 20:45 ` Joseph Myers
2020-04-08 10:15 ` paul zimmermann
0 siblings, 2 replies; 5+ messages in thread
From: Joseph Myers @ 2020-04-07 20:36 UTC (permalink / raw)
To: paul zimmermann; +Cc: libc-alpha
On Mon, 6 Apr 2020, paul zimmermann wrote:
> I disabled the code in math/w_exp10f_compat.c by adding #if 0 ... #endif
> around it, since I don't know how to properly do it.
I'd expect symbol versioning like for expf. So
#if LIBM_SVID_COMPAT && SHLIB_COMPAT (libm, GLIBC_2_0, GLIBC_2_32)
in w_exp10f_compat.c, and math/w_exp10f.c using symbol versioning with
GLIBC_2_32 version, and sysdeps/ieee754/w_expf.c overriding that with a
dummy file, and e_expf.c using a GLIBC_2_32 symbol version, and
math/Versions being updated with the new symbol version, and all the ABI
test baselines being updated as well.
Everything apart from the benchtests will need to go in a single commit
(thus a single patch in the submission), rather than having later patches
fix up problems with earlier ones, on the principle of bisectability.
Benchtests can go in a separate commit.
--
Joseph S. Myers
joseph@codesourcery.com
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2020-04-08 11:35 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-04-06 11:51 faster exp10f code paul zimmermann
2020-04-07 20:36 ` Joseph Myers
2020-04-07 20:45 ` Joseph Myers
2020-04-08 10:15 ` paul zimmermann
2020-04-08 11:35 ` Adhemerval Zanella
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).