[music-dsp] Equalizer

eli+ at gs211.sp.cs.cmu.edu eli+ at gs211.sp.cs.cmu.edu
Fri Nov 23 21:29:26 EST 2001

I've never implemented a graphical EQ, so I'm just kibitzing with the
goal of firming up the FAQ entry here.

Jon Watte wrote:
> Putting 10 filters in series runs the quite significant risk that any 
> inadequacy in the filter design (and there will be inadequacies) can 
> multiply 10-fold, so I'm not so sure that series is "better".

Interesting.  Meaning that numerical error is a larger problem (with
fp math?) than I'd have expected, or that the ideal para-EQ filter
itself is apt to do something inappropriate here?

> Also, my claim was for "cheaper". You can clearly run 20 filters in 
> 5 iterations (using SSE) or 10 iterations (using 3DNow!) when you run 
> them in parallel. Not so when you run them in series.

Does it not work to unroll the loop and stagger the data?  You know,
where say you're doing two filters f0 and f1 in series,

input   x0
output  x2 = f1 x1 = f1 (f0 x0)

// init
x1(0) = f0 x0(0)
// loop
[x1(1) x2(0)] = [f0 f1] [x0(1) x1(0)]   // a length-2 vector computation
[x1(2) x2(1)] = [f0 f1] [x0(2) x1(1)]

> I think this subject, too, should go in the FAQ :-)

Also, those details of how to fix up the parallel filterbank's phase
to avoid cancellation problems.

     Eli Brandt  |  eli+ at cs.cmu.edu  |  http://www.cs.cmu.edu/~eli/

dupswapdrop -- the music-dsp mailing list and website: subscription info,
FAQ, source code archive, list archive, book reviews, dsp links

More information about the music-dsp mailing list