Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]


Groups > linux.kernel > #1194681 > unrolled thread

Re: [PATCH] lib: Make _find_next_bit helper function inline

Started byYury <yury.norov@gmail.com>
First post2015-07-28 23:30 +0200
Last post2015-07-29 22:50 +0200
Articles 5 — 4 participants

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.


Contents

  Re: [PATCH] lib: Make _find_next_bit helper function inline Yury <yury.norov@gmail.com> - 2015-07-28 23:30 +0200
    Re: [PATCH] lib: Make _find_next_bit helper function inline Yury <yury.norov@gmail.com> - 2015-07-28 23:40 +0200
    Re: [PATCH] lib: Make _find_next_bit helper function inline Andrew Morton <akpm@linux-foundation.org> - 2015-07-28 23:50 +0200
      Re: [PATCH] lib: Make _find_next_bit helper function inline Alexey Klimov <klimov.linux@gmail.com> - 2015-07-29 15:40 +0200
        Re: [PATCH] lib: Make _find_next_bit helper function inline Cassidy Burden <cburden@codeaurora.org> - 2015-07-29 22:50 +0200

#1194681 — Re: [PATCH] lib: Make _find_next_bit helper function inline

FromYury <yury.norov@gmail.com>
Date2015-07-28 23:30 +0200
SubjectRe: [PATCH] lib: Make _find_next_bit helper function inline
Message-ID<pRkBJ-4uH-19@gated-at.bofh.it>
On 28.07.2015 22:09, Cassidy Burden wrote:
> I've tested Yury Norov's find_bit reimplementation with the test_find_bit
> module (https://lkml.org/lkml/2015/3/8/141) and measured about 35-40%
> performance degradation on arm64 3.18 run with fixed CPU frequency.
>
> The performance degradation appears to be caused by the
> helper function _find_next_bit. After inlining this function into
> find_next_bit and find_next_zero_bit I get slightly better performance
> than the old implementation:
>
> find_next_zero_bit          find_next_bit
> old      new     inline     old      new     inline
> 26       36      24         24       33      23
> 25       36      24         24       33      23
> 26       36      24         24       33      23
> 25       36      24         24       33      23
> 25       36      24         24       33      23
> 25       37      24         24       33      23
> 25       37      24         24       33      23
> 25       37      24         24       33      23
> 25       36      24         24       33      23
> 25       37      24         24       33      23
>
> Signed-off-by: Cassidy Burden <cburden@codeaurora.org>
> Cc: Alexey Klimov <klimov.linux@gmail.com>
> Cc: David S. Miller <davem@davemloft.net>
> Cc: Daniel Borkmann <dborkman@redhat.com>
> Cc: Hannes Frederic Sowa <hannes@stressinduktion.org>
> Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
> Cc: Mark Salter <msalter@redhat.com>
> Cc: AKASHI Takahiro <takahiro.akashi@linaro.org>
> Cc: Thomas Graf <tgraf@suug.ch>
> Cc: Valentin Rothberg <valentinrothberg@gmail.com>
> Cc: Chris Wilson <chris@chris-wilson.co.uk>
> ---
>   lib/find_bit.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/lib/find_bit.c b/lib/find_bit.c
> index 18072ea..d0e04f9 100644
> --- a/lib/find_bit.c
> +++ b/lib/find_bit.c
> @@ -28,7 +28,7 @@
>    * find_next_zero_bit.  The difference is the "invert" argument, which
>    * is XORed with each fetched word before searching it for one bits.
>    */
> -static unsigned long _find_next_bit(const unsigned long *addr,
> +static inline unsigned long _find_next_bit(const unsigned long *addr,
>   		unsigned long nbits, unsigned long start, unsigned long invert)
>   {
>   	unsigned long tmp;

Hi Cassidi,

At first, I'm really surprised that there's no assembler implementation
of find_bit routines for aarch64. Aarch32 has ones...

I was thinking on inlining the helper, but decided not to do this....

1. Test is not too realistic. https://lkml.org/lkml/2015/2/1/224
The typical usage pattern is to look for a single bit or range of bits.
So in practice nobody calls find_next_bit thousand times.

2. Way more important to fit functions into as less cache lines as
possible. https://lkml.org/lkml/2015/2/12/114
In this case, inlining increases cache lines consumption almost twice...

3. Inlining prevents compiler from some other possible optimizations. It's
probable that in real module compiler will inline callers of _find_next_bit,
and final output will be better. I don't like to point out the compiler how
it should do its work.

Nevertheless, if this is your real case, and inlining helps, I'm OK with it.

But I think, before/after for x86 is needed as well.
And why don't you consider '__always_inline__'? Simple inline is only a 
hint and
guarantees nothing.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [next] | [standalone]


#1194690

FromYury <yury.norov@gmail.com>
Date2015-07-28 23:40 +0200
Message-ID<pRkLo-4Fc-21@gated-at.bofh.it>
In reply to#1194681
On 29.07.2015 00:23, Yury wrote:
> On 28.07.2015 22:09, Cassidy Burden wrote:
>> I've tested Yury Norov's find_bit reimplementation with the 
>> test_find_bit
>> module (https://lkml.org/lkml/2015/3/8/141) and measured about 35-40%
>> performance degradation on arm64 3.18 run with fixed CPU frequency.
>>
>> The performance degradation appears to be caused by the
>> helper function _find_next_bit. After inlining this function into
>> find_next_bit and find_next_zero_bit I get slightly better performance
>> than the old implementation:
>>
>> find_next_zero_bit          find_next_bit
>> old      new     inline     old      new     inline
>> 26       36      24         24       33      23
>> 25       36      24         24       33      23
>> 26       36      24         24       33      23
>> 25       36      24         24       33      23
>> 25       36      24         24       33      23
>> 25       37      24         24       33      23
>> 25       37      24         24       33      23
>> 25       37      24         24       33      23
>> 25       36      24         24       33      23
>> 25       37      24         24       33      23
>>
>> Signed-off-by: Cassidy Burden <cburden@codeaurora.org>
>> Cc: Alexey Klimov <klimov.linux@gmail.com>
>> Cc: David S. Miller <davem@davemloft.net>
>> Cc: Daniel Borkmann <dborkman@redhat.com>
>> Cc: Hannes Frederic Sowa <hannes@stressinduktion.org>
>> Cc: Lai Jiangshan <laijs@cn.fujitsu.com>
>> Cc: Mark Salter <msalter@redhat.com>
>> Cc: AKASHI Takahiro <takahiro.akashi@linaro.org>
>> Cc: Thomas Graf <tgraf@suug.ch>
>> Cc: Valentin Rothberg <valentinrothberg@gmail.com>
>> Cc: Chris Wilson <chris@chris-wilson.co.uk>
>> ---
>>   lib/find_bit.c | 2 +-
>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/lib/find_bit.c b/lib/find_bit.c
>> index 18072ea..d0e04f9 100644
>> --- a/lib/find_bit.c
>> +++ b/lib/find_bit.c
>> @@ -28,7 +28,7 @@
>>    * find_next_zero_bit.  The difference is the "invert" argument, which
>>    * is XORed with each fetched word before searching it for one bits.
>>    */
>> -static unsigned long _find_next_bit(const unsigned long *addr,
>> +static inline unsigned long _find_next_bit(const unsigned long *addr,
>>           unsigned long nbits, unsigned long start, unsigned long 
>> invert)
>>   {
>>       unsigned long tmp;
>
> Hi Cassidi,
>
> At first, I'm really surprised that there's no assembler implementation
> of find_bit routines for aarch64. Aarch32 has ones...
>
> I was thinking on inlining the helper, but decided not to do this....
>
> 1. Test is not too realistic. https://lkml.org/lkml/2015/2/1/224
> The typical usage pattern is to look for a single bit or range of bits.
> So in practice nobody calls find_next_bit thousand times.
>
> 2. Way more important to fit functions into as less cache lines as
> possible. https://lkml.org/lkml/2015/2/12/114
> In this case, inlining increases cache lines consumption almost twice...
>
> 3. Inlining prevents compiler from some other possible optimizations. 
> It's
> probable that in real module compiler will inline callers of 
> _find_next_bit,
> and final output will be better. I don't like to point out the 
> compiler how
> it should do its work.
>
> Nevertheless, if this is your real case, and inlining helps, I'm OK 
> with it.
>
> But I think, before/after for x86 is needed as well.
> And why don't you consider '__always_inline__'? Simple inline is only 
> a hint and
> guarantees nothing.

(Sorry for typo in your name. Call me Yuri next time.)

Adding Rasmus and George to CC

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]


#1194698

FromAndrew Morton <akpm@linux-foundation.org>
Date2015-07-28 23:50 +0200
Message-ID<pRkV4-4Qy-7@gated-at.bofh.it>
In reply to#1194681
On Wed, 29 Jul 2015 00:23:18 +0300 Yury <yury.norov@gmail.com> wrote:

> But I think, before/after for x86 is needed as well.

That would be nice.

> And why don't you consider '__always_inline__'? Simple inline is only a 
> hint and
> guarantees nothing.

Yup.  My x86_64 compiler just ignores the "inline".  When I use
__always_inline, find_bit.o's text goes from 776 bytes to 863. 
Hopefully we get something in return for that bloat!


Also, if _find_next_bit() benefits from this then _find_next_bit_le()
will presumably also benefit.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]


#1195192

FromAlexey Klimov <klimov.linux@gmail.com>
Date2015-07-29 15:40 +0200
Message-ID<pRzKs-15P-61@gated-at.bofh.it>
In reply to#1194698
On Вт., 2015-07-28 at 14:45 -0700, Andrew Morton wrote:
> On Wed, 29 Jul 2015 00:23:18 +0300 Yury <yury.norov@gmail.com> wrote:
> 
> > But I think, before/after for x86 is needed as well.
> 
> That would be nice.
> 
> > And why don't you consider '__always_inline__'? Simple inline is only a 
> > hint and
> > guarantees nothing.
> 
> Yup.  My x86_64 compiler just ignores the "inline".  When I use
> __always_inline, find_bit.o's text goes from 776 bytes to 863. 
> Hopefully we get something in return for that bloat!

On my x86_64 (core-i5 something, with disabled cpufreq) i got following
numbers:
find_next_zero_bit
old	new	__always_inline
20	21	22	
20	21	22
20	22	23
21	21	22
21	21	23
20	21	22
20	21	23
21	22	23
20	22	22
21	21	22

find_next_bit
old	new	__always_inline
19	21	24
19	22	24
19	22	24
19	21	24
20	22	24
19	21	23
19	21	23
20	21	24
19	22	24
19	21	24

I will re-check on another machine. It's really interesting if
__always_inline makes things better for aarch64 and worse for x86_64. It
will be nice if someone will check it on x86_64 too.

Best regards,
Alexey Klimov.

> Also, if _find_next_bit() benefits from this then _find_next_bit_le()
> will presumably also benefit.
> 
> 


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [next] | [standalone]


#1195492

FromCassidy Burden <cburden@codeaurora.org>
Date2015-07-29 22:50 +0200
Message-ID<pRGsy-2ir-11@gated-at.bofh.it>
In reply to#1195192
I changed the test module to now set the entire array to all 0/1s and
only flip a few bits. There appears to be a performance benefit, but
it's only 2-3% better (if that). If the main benefit of the original
patch was to save space then inlining definitely doesn't seem worth the
small gains in real use cases.

find_next_zero_bit (us)
old      new     inline
14440    17080   17086
4779     5181    5069
10844    12720   12746
9642     11312   11253
3858     3818    3668
10540    12349   12307
12470    14716   14697
5403     6002    5942
2282     1820    1418
13632    16056   15998
11048    13019   13030
6025     6790    6706
13255    15586   15605
3038     2744    2539
10353    12219   12239
10498    12251   12322
14767    17452   17454
12785    15048   15052
1655     1034    691
9924     11611   11558

find_next_bit (us)
old      new     inline
8535     9936    9667
14666    17372   16880
2315     1799    1355
6578     9092    8806
6548     7558    7274
9448     11213   10821
3467     3497    3449
2719     3079    2911
6115     7989    7796
13582    16113   15643
4643     4946    4766
3406     3728    3536
7118     9045    8805
3174     3011    2701
13300    16780   16252
14285    16848   16330
11583    13669   13207
13063    15455   14989
12661    14955   14500
12068    14166   13790

On 7/29/2015 6:30 AM, Alexey Klimov wrote:
> I will re-check on another machine. It's really interesting if
> __always_inline makes things better for aarch64 and worse for x86_64. It
> will be nice if someone will check it on x86_64 too.

Very odd, this may be related to the other compiler optimizations Yuri
mentioned?
-- 
The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[toc] | [prev] | [standalone]


Back to top | Article view | linux.kernel


csiph-web