On Oct 9, 5:38=A0am, cs_post...@[EMAIL PROTECTED]
wrote:
> On Oct 7, 10:13 am, move <liubeny...@[EMAIL PROTECTED]
> wrote:
>
> > I am using a Differential delay=3D50 and Number of sections=3D5 CIC
fil=
ter
> > and cascading a ciccomp filter which provide an extra 2 decimator.
>
> I don't think you mean your decimation ratio is 50, not your
> differential delay.
>
> One solution is to come out of your CIC at a higher rate, and
> decimator by 4 or 8 in your compensation filter. =A0 This means that you
> are using a narrower slice at the center of the CIC, so there's less
> droop to compensate. =A0In fact your may be able to skip compensation
> all together. =A0Then what was your compensation filter just becomes a
> decimate by four or eight polyphase FIR lowpass.
>
> > will take 404 taps to implement using equiripple method, the taps is
> > too long for input data block which is only 200 samples a frame, so
> > before next frame is coming, the filter is still full of unprocessed
> > last frame data, this will cause the output of nextframe filter in-
> > accurate.
>
> Shouldn't you be joining the frames together, processing them as a
> continuous stream, and then reframing them again at the output? =A0If
> the next frame hasn't arrived, just freeze your filter logic in place
> and don't give it another clock until you have data from the next
> frame to start feeding in.
Thanks!
about the decimation rate : when I design using Xilinx Ip Core, the
decimation is made to be an integral part of CIC filter, i.e the
decimation and defferential delay are made together, if M=3D50 and D=3D50,
the decimator should be placed between the comb and integer part, and
then synthesis using decimation=3D50 and D=3D1.
I am wonderring whether i could reduce the taps of CICcomp filter is
that the ciccomp filter design in matlab could not be implemented in a
multistage manner. ('cause the shape in the passband is not flat and
inv-sinc), when the taps of filter exceed the number of input, i
reckon the filter not a success one or padding zeros is needed to make
the input length N+L-1. but once the situation is time critial (in a
frame), padding zeros is not valid.
|
