Path: csiph.com!usenet.pasdenom.info!weretis.net!feeder1.news.weretis.net!feeder.erje.net!xlned.com!feeder7.xlned.com!news2.euro.net!postnews2.euro.net!news.wanadoo.nl!not-for-mail
From: mhx@iae.nl (Marcel Hendrix)
Subject: Re: SSE2
Newsgroups: comp.lang.forth
Message-ID: <81078807968435@frunobulax.edu>
Date: Sat, 28 Jul 2012 15:28:54 +0200
References: <2012Jul27.165605@mips.complang.tuwien.ac.at>
X-Newsreader: iForth 2.0 console (October 21, 2006)
Lines: 40
Organization: Wanadoo
NNTP-Posting-Date: 28 Jul 2012 13:28:27 GMT
NNTP-Posting-Host: s529d937f.adsl.wanadoo.nl
X-Trace: 1343482107 dr6.euro.net 221 82.157.147.127:64505
X-Complaints-To: abuse@wanadoo.nl
Xref: csiph.com comp.lang.forth:14481

anton@mips.complang.tuwien.ac.at (Anton Ertl) writes Re: SSE2

> mhx@iae.nl (Marcel Hendrix) writes:
>>What I did is derived from the MiniBLAS sources. As SSE2 operates on 
>>4 doubles at a time, speedups of around 4 are suggesting themselves. 
>>However, I can find no trace of this. An obvious reason could be that
>>memory throughput can not keep up with the FP units. Strange, as one 
>>would expect this hardware problem to be fixed by now.

> If the problem in your benchmark is that it is memory bandwidth
> limited, no, that problem is not "fixed". 

See my answer to Paul. The problem was that the S/DDOT code needs
large vectors to become effective. For small sizes it does almost
nothing. 

[,,]

> Hmm, if the problem is memory bandwidth, I would expect all variants
> to have the same performance (unless you use additional indirection
> vectors or varying memory layouts).

That's a good argument.

[..]
> My _guess_ is that your compiler produces some stack memory stores for
> some of the variants, with some stack memory fetches soon after, and
> these cost quie a bit of performance.  At least they used to.

I would be interested to read more about this. What is the problem here,
exactly? Is a stackframe better than push/pop?

>                                                                You
> don't have a data-near-code problem, do you?

No, I allways check for these, but I have not come across anything 
like it since quite a long time.

-marcel