From: Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>
To: Wilco Dijkstra <Wilco.Dijkstra@arm.com>,
'GNU C Library' <libc-alpha@sourceware.org>
Subject: Re: [PATCH] math: remove exp10 wrappers
Date: Thu, 11 Jan 2024 09:35:01 -0300 [thread overview]
Message-ID: <b6df2de6-51d0-441e-8c91-9fb5b3edefc1@linaro.org> (raw)
In-Reply-To: <PAWPR08MB89825D30F903B039B7AF492083692@PAWPR08MB8982.eurprd08.prod.outlook.com>
On 10/01/24 18:35, Wilco Dijkstra wrote:
> Hi Adhemerval,
>
>> As a side node, I am checking if it is possible to remove the i686 version
>> in favor of the generic one and it seems that the double_t usage on generic
>> implementation does not guarantee either a faster or conformant implementation
>> (i386 uses FLT_EVAL_METHOD=2 and the resulting exp10 shows a lot precision
>> issues with math tests). Does it really need to use double_t?
>
> Yes it would be great to remove target specific math implementations. With IA64
> removed I think the only ones left are x86 and m68k. Both would be better off using
> the generic math code.
>
> We added double_t to use wider internal precision without storing to memory all the
> time. In principle it should be faster overall. Exp10 uses double_t like exp and exp2,
> so I'm surprised it doesn't work. I do notice both exp/exp2 do ki = asuint64 (kd);
> while exp10 just does ki = kd, but it's not clear that would make a difference on x86.
>
The patch below fixes the numerical stability I found on i686:
diff --git a/sysdeps/ieee754/dbl-64/e_exp10.c b/sysdeps/ieee754/dbl-64/e_exp10.c
index 225fc74c4c..f961d58398 100644
--- a/sysdeps/ieee754/dbl-64/e_exp10.c
+++ b/sysdeps/ieee754/dbl-64/e_exp10.c
@@ -99,7 +99,7 @@ __exp10 (double x)
/* Reduce x: z = x * N / log10(2), k = round(z). */
double_t z = __exp_data.invlog10_2N * x;
- double_t kd;
+ double kd;
int64_t ki;
#if TOINT_INTRINSICS
kd = roundtoint (z);
And I think we haven't see it because all ABIs with FLT_EVAL_METHOD=2 uses
assembly implementations for exp/exp2/exp10 (this is the case for i386
and m68k).
Furthermore, at least with i686 the use of double_t also does not show
better performance either. The current sysdeps/i386/fpu/e_exp10.S
on a recent CPU (Ryzen 9 5900X) with gcc version 13.2.1:
$ ./benchtests/bench-exp10
"exp10": {
"": {
"duration": 3.71309e+09,
"iterations": 5.8534e+07,
"max": 116.55,
"min": 62.974,
"mean": 63.4348
}
The fixed version using double_t:
$ ./benchtests/bench-exp10
"exp10": {
"": {
"duration": 3.71002e+09,
"iterations": 6.5653e+07,
"max": 130.499,
"min": 50.282,
"mean": 56.5096
}
And just using double instead of double_t:
$ ./benchtests/bench-exp10
"exp10": {
"": {
"duration": 3.71803e+09,
"iterations": 7.3563e+07,
"max": 123.062,
"min": 43.549,
"mean": 50.5421
}
}
So I am not sure if double_t is really helpful here.
next prev parent reply other threads:[~2024-01-11 12:35 UTC|newest]
Thread overview: 12+ messages / expand[flat|nested] mbox.gz Atom feed top
2024-01-10 16:03 Wilco Dijkstra
2024-01-10 18:48 ` Adhemerval Zanella Netto
2024-01-10 21:35 ` Wilco Dijkstra
2024-01-11 2:11 ` Joseph Myers
2024-01-11 9:06 ` Andreas Schwab
2024-01-11 12:35 ` Adhemerval Zanella Netto [this message]
2024-01-16 15:41 ` Wilco Dijkstra
2024-01-16 16:08 ` Szabolcs Nagy
2024-01-16 16:11 ` Adhemerval Zanella Netto
2024-01-18 14:26 ` Wilco Dijkstra
2024-01-18 15:30 ` Szabolcs Nagy
2024-01-11 21:46 ` Andreas K. Huettel
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=b6df2de6-51d0-441e-8c91-9fb5b3edefc1@linaro.org \
--to=adhemerval.zanella@linaro.org \
--cc=Wilco.Dijkstra@arm.com \
--cc=libc-alpha@sourceware.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
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).