FW: Incorrect samplerate from RTL-SDR

Jiří Pinkava j-pi at seznam.cz
Fri Jan 10 09:52:07 UTC 2014

Dne 10.1.2014 05:23, Richard Koch napsal(a):
> JP,
> Thanks for your reply and explanation.
> I understand the cutting the coefficient table in half since it
> repeats in a backwards manner, but what I don't understand
> is why half the table is in 8 bit and the other in 12 bit?
8 / 12 bit is used inside the dongle and is how the hardware is designed.
It reduces the cost of hardware. You can think of it in the way
the coefficients rea all 12 bit, but for the first 8 coefficients the
top 4 bits must
be always zero.
This is only the bite width of coefficients not the actual width of
result. Unfortunately I'm not able to explain details about FIR and
hardware design here.

I'm not sure if I explained your question, try reformulate and extend it
if not.

> Also I now understand that there is not enough room for extra
> coefficients in the rtl2832 for lower bandwidth than 2048Khz.
> That is unfortunate if you want lower bandwidth as it forces you
> to sample at 2048 or greater and perform your own FIR and
> decimation consuming CPU cycles on your host. On an armv4
> SBC that is very taxing :^)
Working with such hardware is always a challenge, but the tweaked result
is often beautifull.
Good luck,
> -Rick
>> Date: Fri, 10 Jan 2014 01:50:51 +0100
>> From: j-pi at seznam.cz
>> To: n1gp at hotmail.com; osmocom-sdr at lists.osmocom.org
>> Subject: Re: Incorrect samplerate from RTL-SDR
>> CC: steve at steve-m.de
>> You cannot fit internal FIR filter in SDR to such narrow bandwidth, it
>> does not have enough coefficients.
>> You can adapt it only to wider band width (some small improvement can be
>> done for
>> sample rates above 2048 kHz). Select proper values is a bit tricky and
>> specialised software
>> is almost unavoidable. You can get good hint from FIR filter designer in
>> GNU Radio, but
>> this software is not designed for this task and the work is cumbersome.
>> The 8 / 12 bit means the coefficients  can be only in range form -128 to
>> 127 / -2048 to 2047.
>> The usual FIR coefficients absolute value decreases from center to the
>> side eg:
>> 1, -3, 5, 10, 5, -3, 1,
>> so this just reduces complexity of the device.
>> For lower bandwidth you need more cofficients (more samples because the
>> wave is wider).
>> JP
>> Dne 9.1.2014 23:53, Richard Koch napsal(a):
>>> Hi Steve,
>>> In regards to your post about the anti-aliasing filter
>>> you referred to, it appears that it is getting set in
>>> librtlsdr.c, rtlsdr_init_baseband()
>>> If one wanted to change that to optimize the anti-aliasing
>>> to a lower bandwidth, say 1500 Khz instead of the 2048 Khz
>>> How would you change the fir_coeff[] table entries to achieve
>>> this? The 8 bit entries and 12 bit entries are confusing to me.
>>> ==== orig post ====
>>>> I have seen other reports on the mailing list, saying that sample rates
>>>> between 300 kS/s and 900 kS/s don't work. If this applies to all RTL
>>>> devices, then maybe it should be documented and the library can return
>>>> an error code for such sample rates. Otherwise people will keep walking
>>>> into this trap.
>>> We could return some sort of error in those cases, but such low rates
>>> (< 1MS/s) aren't recommended in general. First of all, the
>>> anti-aliasing filter we're using has a fixed bandwidth of 2 MHz, and
>>> although the coefficients can be changed, I doubt you can get a nice
>>> filter for lower bandwidths with such a low order. And then the ADCs
>>> only have 8 bits resolution, so you want to improve that by decimating,
>>> and also profit from decimation gain.
>>> Even Realtek uses 2.048 MS/s for FM reception in their original Windows
>>> software. 		 	   		  

More information about the osmocom-sdr mailing list