Re: [Discuss-gnuradio] resample ratio for fractional_resampler

[Marcus Müller]

Hi Andy,

The fallacy I had in my own head was that the MMSE resampler also 
filters sufficiently – I think that is not true; it's just a resampler, 
you need to "select" (as in: filter everything else away) the 
appropriate bandwidth yourself before.

Therefore: Yang, you'll need to filter before to restrict your signal to 
a Nyquist bandwidth that the resampled signal can represent without 
ambiguities (anti-aliasing), then have the flexible MMSE resampler, 
then, if necessary, suppress images. That's at least one times the 
effort of the 1/64-bandwidth FIR from my last mail that you *must* do, 
if your signal is not already sufficiently-bandlimited before.

That leads me to the conclusion that you'd definitely want to first 
decimate with a FIR and then flexibly interpolate/resample.

You can, in fact, combine both, and then you (essentially) get working 
principle of the "Polyphase arbitrary resampler": Build a polyphase 
filterbank from a decimation filter (that's the "Decimating FIR" 
aspect), and go through the phases in a way that approximates the 
sub-input-sample delay you want at the output (that's the 
MMSE-filterbank-alike idea about that).

So, Yang, could you elaborate about what you do before (and after) the 
resampler? What does your input signal's spectrum look like?

Best regards,
Marcus


On 2017-10-24 14:22, Andy Walls wrote:
> On Tue, 2017-10-24 at 08:16 -0400, Andy Walls wrote:
> > Hi Marcus:
> >
> > Nice insights and plots.
> >
> > > From:   Marcus Müller
> > > Date:   Mon, 23 Oct 2017 20:10:23 +0200
> > [snip]
> > >  From the above figure, I'd say: go for 256 taps; 50 dB of anti-
> > > aliasing should probably suffice.
> > > On my machine, the 256 tap FIR filter with decimation=64 ran about
> > > 20% faster than the fractional resampler; comparing what they do to
> > > white noise, I'd say that whilst the spectral shape of the 256 tap
> > > FIR is not perfect, the MMSE is pretty much unusable, I guess:
> > I'd just like to interject here that GNURadio's MMSE filterbank is
> > only
> > MMSE for input spectral components in the range [-sample_rate/4,
> > sample_rate/4].
> >
> > Also it appears that the mmse/fractional_resampler blocks don't
> > perform
> > any explicit anti-alias filtering.  So that's another reason to have
> > a
> > band-limiting filter out in front of it.
> Oopsie, and maybe one after it too, since we're talking about
> decimation.
>
> -Andy

smime.p7s
Description: S/MIME Cryptographic Signature