Sean, I guess there was a missing normalization of the pulse.
Please add the line p=p/sum(p)*Q/2.0; just after the definition of the p pulse p = numpy.ones(L*Q) In addition, make sure you increase the Q (samples per symbol) considerably, ie make it 16 or more. The reason is that h=25/3 results in a BIG bandwidth for this constellation so it has to be appropriately samples in order to be represented accurately. let me know how it goes Achilleas On Thu, Sep 10, 2015 at 6:09 PM, Nowlan, Sean <sean.now...@gtri.gatech.edu> wrote: > Thanks, Achilleas, Jan, Jeff for your responses. I think I understand the > format now, and I managed to build the gr-trellis documentation, which does > spell out answers for a lot of what I asked. > > > > I’ve abandoned making my own FSM since I can just use the CPM utilities in > gr-trellis. However, I’m getting unexpected BER and PER when running > gr-trellis/examples/python/test_cpm.py with my CPFSK parameters, h = 25/3 > and L=1. > > > > Diff of test_cpm.py with my simple modifications: > > https://gist.github.com/nowls/5b2bc2aada87dc197e3f > > > > $ python test_cpm.py > > Using Volk machine: avx_64_mmx_orc > > 100 16868 100 0.16868 1.0 > > 200 33852 200 0.16926 1.0 > > 300 50786 300 0.169286666667 1.0 > > 400 67667 400 0.1691675 1.0 > > […] > > > > As you can see, PER is 100% and BER is pretty bad for 10 dB SNR. The > construction of the phase response, q, looks correct to me. Any idea what > might be going wrong? > > > > Thanks, > > Sean > > > > *From:* discuss-gnuradio-bounces+sean.nowlan=gtri.gatech....@gnu.org > [mailto:discuss-gnuradio-bounces+sean.nowlan=gtri.gatech....@gnu.org] *On > Behalf Of *Achilleas Anastasopoulos > *Sent:* Thursday, September 10, 2015 1:28 PM > *To:* Discuss-gnuradio@gnu.org > *Subject:* Re: [Discuss-gnuradio] Demodulating CPFSK with Viterbi > algorithm using gr-trellis > > > > Sean, > > I just wanted to add to the previous answer: > > Indeed gr-trellis (and the fsm class specifically) has a built in > constructor for generating the FSM corresponding to ANY CPM scheme. > > It only requires three parameters as explained in the comments > > in the file fsm.cc > > The constructed FSM is NOT based on Proakis decomposition, but on the more > elegant decomposition by: > "A decomposition approach to CPM", IEEE Trans. Info Theory, March 1988 > See also my own notes at > http://www.eecs.umich.edu/~anastas/docs/cpm.pdf > > for all the details. > > In this case for h=K/P, it does not matter whether K is even or odd, the > number of states is always = M^{L-1} * P (where M is the cardinality of the > input alphabet), > > so in your case you'll have a trellis with 3 states as the previous email > suggested. > > You may also find useful the python utility > > make_cpm_signals(K,P,M,L,q,frac) > > that can be found in gr-trellis/python > > that generates the signal space for you. > > (please see the above references for more info). > > Finally, you can take a look at the example "test_cpm.py" > > in the gr-trellis/examples/python > > and make appropriate changes for your CPFSK signal. > > > ------------------------ > > Now you can always drop all the above and design your FSM and signals from > scratch. You can create the FSM from a file (this is explained in the > gr-trellis documentaion--which is generated from > gr-trellis/doc/gr-trellis.xml). > > The ouput symbol mapping is defined EXACTLY as you suggested in your email: > > every line corresponds to the current state and every column to an input. > > The entry is a number from 0 to I-1 (where I is the input cardinality). > > How this index is translated to an actual signal IS NOT PART OF THE FSM > DEFINITION but part of the modulation definition as explained very > extensively in the gr-trellis documentation. > > > > let me know if you have any further questions, > > Achilleas > > > > > > >
_______________________________________________ Discuss-gnuradio mailing list Discuss-gnuradio@gnu.org https://lists.gnu.org/mailman/listinfo/discuss-gnuradio
