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: Andi Kleen <andi at firstfloor dot org>
- To: OndÅej BÃlka <neleai at seznam dot cz>
- Cc: Andi Kleen <andi at firstfloor dot org>, Carlos O'Donell <carlos at redhat dot com>, Dmitrieva Liubov <liubov dot dmitrieva at gmail dot com>, "H.J. Lu" <hjl dot tools at gmail dot com>, GNU C Library <libc-alpha at sourceware dot org>
- Date: Fri, 21 Jun 2013 15:06:19 +0200
- Subject: Re: [PATCH] Don't use SSE4_2 instructions on Intel Silvermont Micro Architecture.
- References: <CAMe9rOqT31AFq1S3V0Krh2CZnHu=FiyXqhg840fimRtfU4_hXQ at mail dot gmail dot com> <20130618064910 dot GA19972 at domone dot kolej dot mff dot cuni dot cz> <CAHjhQ90Fc0kdZfQrUwLwpKbz2va4X9rzf1EkGD-s-RH-iF7guQ at mail dot gmail dot com> <CAHjhQ92qfjdKZthqAwxCVuCnLqDr2stdEbQpne5rKhzJPeN_cQ at mail dot gmail dot com> <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>
> To see if this is a case I added Andi. Andi, could you browse sources
> and tell if you think that benchtests are adequate to measure
> performance?
It's hard to say in general and just from review.
Normally I would recommend to not run the tests in a tight loop,
but have some cache / branch prediction thrashing code in between
to measure "cache cold" behaviour too.
One objective approach to determine how good they are would be
to assemble benchmark tests on a set of non micro applications that are known
to be sensitive to memcpy/etc. performance (e.g. according to
perf sampling). Then do a set of results for them plus run the benchmarks.
Change the memcpy et.al. to make it slower (for example or faster if you have
a good way). Rerun the applications. Rerun the benchmarks.
Then see how the two sets of benchmark results correlate.
This would be significant work of course.
-Andi