Re: [linux-audio-dev] float a/d (was Linux support for IEEE1394 mLAN?)

New Message Reply About this list Date view Thread view Subject view Author view Other groups

Subject: Re: [linux-audio-dev] float a/d (was Linux support for IEEE1394 mLAN?)
From: Jussi Laako (jussi_AT_jlaako.pp.fi)
Date: Wed Oct 04 2000 - 21:10:14 EEST


Benno Senoner wrote:
>
> That's not always true, unless you access large arrays of FP data very
> frequently.

What is large? I think 1024 point FFT as small, I'm usually doing 4096 -
16384 point FFT's in realtime. And 128 - 512 point FIR filters, cascaded.
I'm also doing matrix operations on matrixes of size 1024x1024 and
1024x1024x128.

> Plus think about the fact that if you use small fragment sizes, a big
> part of the involved FP variables will be chaches, thus the
> single-precison FP advantage partially vanishes in this case.

This is one of main reasons why single FP is faster than double FP. You can
just fit twice the data into cache when using single FP.

> I have no numbers yet, but EVO will be a good app to test this.

I've been testing this alot.

> (in fact I have a pure sampleplayback engine (resampling and mixing), and
> when switching from doubles to floats the CPU usage drops only minimally,
> but I assure you that heavy duty stress tests (float vs double) will
> follow within the next few months).

Good mix of float and double is to code in assembler and keep data in FP
stack (registers) and get it out as float.

> interconnected by 100Mbit/1Gbit ethernet / Myrinet cards.

I've done some testing with 100M clusters, nice... ;-)

 - Jussi Laako

-- 
PGP key fingerprint: 161D 6FED 6A92 39E2 EB5B  39DD A4DE 63EB C216 1E4B
Available at: ldap://certserver.pgp.com, http://keys.pgp.com:11371


New Message Reply About this list Date view Thread view Subject view Author view Other groups

This archive was generated by hypermail 2b28 : Wed Oct 04 2000 - 21:58:24 EEST