This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] Don't use SSE4_2 instructions on Intel Silvermont Micro Architecture.\
- From: "Carlos O'Donell" <carlos at redhat dot com>
- To: OndÅej BÃlka <neleai at seznam dot cz>, Siddhesh Poyarekar <siddhesh dot poyarekar at gmail dot com>
- Cc: Liubov Dmitrieva <liubov dot dmitrieva at gmail dot com>, Andi Kleen <andi at firstfloor dot org>, "H.J. Lu" <hjl dot tools at gmail dot com>, GNU C Library <libc-alpha at sourceware dot org>
- Date: Sun, 30 Jun 2013 16:26:11 -0400
- Subject: Re: [PATCH] Don't use SSE4_2 instructions on Intel Silvermont Micro Architecture.\
- References: <51C23583 dot 1070307 at redhat dot com> <CAHjhQ93vWnCiVVU9MPoGptjQtn2J2PCDT2B7ZfXiKt+Cv_Rh_w at mail dot gmail dot com> <51C307A5 dot 7030608 at redhat dot com> <20130620151711 dot GA4891 at domone dot kolej dot mff dot cuni dot cz> <51C317AA dot 6080502 at redhat dot com> <20130621012427 dot GA4574 at domone dot kolej dot mff dot cuni dot cz> <CAAHN_R1HXyy0i25rtYKJ4Zox5u0R57xKbZDq=ZNf0BVm=7biMw at mail dot gmail dot com> <20130621135110 dot GB7973 at domone dot kolej dot mff dot cuni dot cz> <CAHjhQ921kXhi3hfqkHW_5pdYY2QYf6pzQ8OLondc6JJjj++4kQ at mail dot gmail dot com> <51CC602F dot 1010406 at redhat dot com> <20130630195200 dot GA5087 at domone dot homenet dot telecomitalia dot it>
On 06/30/2013 03:52 PM, OndÅej BÃlka wrote:
> On Thu, Jun 27, 2013 at 11:54:23AM -0400, Carlos O'Donell wrote:
>> On 06/27/2013 03:24 AM, Liubov Dmitrieva wrote:
>>> I think for this particular patch we don't need super accurate
>>> benchmarks to see that it is better because we talk not about 20-60%
>>> of boost but about several times asymptotically boost as current
>>> benchmarks showed. It was a server machine, nobody runs Firefox there.
>>
>> Agreed, but we still need some kind of reproducible result that shows
>> your patch improved performance. I'm not happy with performance patches
>> going into glibc without some proof that they made things better.
>>
> You said proof but we are not in proof stage yet. We are not in claim
> stage yet. As these "benchmarks" are like mechr one please explain with
> following code questions below:
>
> for (i = 0; i < 32; ++i)
> {
> HP_TIMING_NOW (start);
> CALL (impl, s, c, n);
> HP_TIMING_NOW (stop);
> HP_TIMING_BEST (best_time, start, stop);
> }
>
> 1. You use only 32 element sample. Can you be sure that this sample is
> big enough to be relevant?
>
> 2. You take minimum of these samples. Please explain how this is related
> to real performance.
>
> 3. You call this code in loop with same argmuments. Please explain why
> real world usage cases are close enough that we can observe same
> behaviour in real world.
>
> Unless you can satisfactory answer these questions you did not prove
> anything about performance only got some numbers that are loosely
> related to it.
Thanks for the review. Could you please start a new thread and CC
Siddhehs whose the most well versed in this code? Your questions
are valid questions, and I hope that their discussion leads to a
better microbenchmark.
Cheers,
Carlos.