Forgot to copy this to the list... > Begin forwarded message: > > From: Steven Franke <[email protected]> > Subject: Re: libjt issue > Date: September 27, 2015 at 7:33:33 PM CDT > To: Joe Taylor <[email protected]> > > Hi Joe - > > Nice job fine-tuning the erasure probabilities! I didn’t think that we’d get > past 800 with erasures alone. With ntrials=10000 it looks like we are now > even with KV. > > - Do you think that it’d be worth implementing an 8x8 erasure probability > matrix? (i.e. carve the metric plane into an 8x8 grid). That might make sense > if you happen to have a good scheme for deriving the probabilities directly > from the histogram data in fort.40. > > - There is still something missing from the Makefile, as it doesn’t seem to > compile encode_rs. This would be easy to fix - but there’s another problem, > caused by my limited knowledge of C, that might make us want to just drop > back to the un-modified Karn routines… > > I originally modified KA9Q’s encode_rs so that the syndrome vector was > calculated only once, on the first call. I changed the syndrome vector's > declaration to be static so that the syndromes, once calculated, would be > remembered on subsequent calls. This works great for the current > erasures-only scheme. Now, when we try to insert mr2sym’s instead of > erasures, we are modifying the un-erased symbols, so it is necessary to > re-calculate the syndromes. I was trying to do that by setting calc_syn = 1 > on calls to decode_rs after whenever an mr2sym had been inserted. I think > that this is incompatible with declaring the syndrome array to be static - it > asks for the array to be modified after it has been initialized. > > I think this explains why, no matter what I did, adding mr2sym insertion was > degrading performance. The easiest fix in the short term might be to revert > to the original KA9Q routines so that we can play with mr2sym insertion. > > Steve k9an >> On Sep 27, 2015, at 2:46 PM, Joe Taylor <[email protected]> wrote: >> >> Steve -- >> >> I moved your modified sources for the _rs_ routines into .../trunk/sfrsd, >> merged sfrsd3.c back into sfrsd2.c, and updated Makefile accordingly. I >> think it should be OK now. As of r5935 I am getting 810/1000 for 10000 >> trials and 874/1000 for 100000 trials. I hope I haven't done something >> stupid that makes these numbers invalid! >> >> I'm now using a heuristic adjustment to the computed "nsoft" when workdat[i] >> fails to match rxdat[i], but *does* match rxdat2[i]. Probably there are >> better ways to do this ... >> >> -- Joe >
------------------------------------------------------------------------------ _______________________________________________ wsjt-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/wsjt-devel
