This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] Improves __ieee754_exp() performance by 2x-5x on sparc/x86.
- From: Joseph Myers <joseph at codesourcery dot com>
- To: Patrick McGehearty <patrick dot mcgehearty at oracle dot com>
- Cc: <libc-alpha at sourceware dot org>
- Date: Wed, 3 Jan 2018 20:57:01 +0000
- Subject: Re: [PATCH] Improves __ieee754_exp() performance by 2x-5x on sparc/x86.
- Authentication-results: sourceware.org; auth=none
- References: <1515007683-3161-1-git-send-email-patrick.mcgehearty@oracle.com>
On Wed, 3 Jan 2018, Patrick McGehearty wrote:
> Using the glibc perf tests on sparc,
> sparc (nsec) x86 (nsec)
> old new old new
> max 17629 395 5173 144
> min 399 54 15 13
> mean 5317 200 1349 23
These are exactly the same figures as in the previous version that was
reverted. Are you *sure* they are the correct results of retesting with
this patch version? Remember, the patch submission message must be
correct for the patch version being submitted and there must be an
unambiguous division between the content intended for the commit message
(describing the current patch version) and any other content such as
descriptions of differences from previous patch versions (which don't go
in the commit message).
I'd suggest following the conventions used by the Linux kernel, of the
description of the current patch version for the commit message going
*first*, then a line "---", then any other text not for the commit message
(such as information about differences from previous patch versions), then
the diff itself.
> In addition, because ieee754_exp() is used by other routines, cexp()
> showed test results with very small imaginary input values where the
> imaginary portion of the result was off by 3 ulp when in upward
> rounding mode, but not in the other rounding modes. For x86, tgamma
> showed a few values where the ulp increased to 6 (max ulp for tgamma
> is 5). Sparc tgamma did not show these failures. I presume the tgamma
> differences are due to compiler optimization differences within the
> gamma function.The gamma function is known to be difficult to compute
> accurately.
Is this still accurate for the current patch version?
If you tested on a platform requiring ulps updates, you should include the
updates from "make regen-ulps" as part of the patch, and explicitly
confirm that, given the patch applied including those ulps updates, all
the math/ tests PASS on platforms that you specify in the patch submission
message.
--
Joseph S. Myers
joseph@codesourcery.com