Joe - If you get a chance, it might be interesting to plot decodes vs ntrials for the sfrsd3.c that I committed just now. Bottom line is that I eventually found a setting that does a significant number of mr2sym insertions and “does no harm” at ntrials=10000, i.e. the number of decodes is not degraded. However, I think that the decodes vs ntrials characteristic may have a slightly different shape, because I see more decodes and smaller execution times at small values of ntrials (e.g. 100, 1000 and 2000).
I did try nesting the mr2sym insertion inside the erasure threshold loop, but I eventually ended up just putting the mr2sym insertions after the erasure insertions. They will occasionally overlap - but so what... Either I’m doing something wrong, or the mr2sym’s just don’t have much to add, at least for the idealized signals that we’re decoding. The story may well be different with real signals. If you want to play, I suggest using the 0.4 factor that is currently in front of the mr2sym threshold as a knob. Setting that factor to 0.0 should yield results equivalent to our current sfrsd2.c. Steve k9an > On Sep 29, 2015, at 2:16 PM, Joe Taylor <j...@princeton.edu> wrote: > > Hi all, > > Some further information on the new Reed Solomon decoder. > > A simple timing measurement shows that for test program "rsdtest" 97% of > the execution time is spent in the Berlekamp-Massey > errors-and-erasures decoder. > > This is good. It means that if/when we decide more speed is desirable, > it should be easy to get it. For an N-CPU machines, N-1 parallel > threads can be used to execute the randomized-erasures loop. > > Extending the number of trials from 10^4 to 10^5 increases the number of > decodes from 837 to 880. Still no false decodes. > > I posted a plot here > http://physics.princeton.edu/pulsar/K1JT/decodes_vs_ntrials.pdf > showing the cumulative number of decodes (out of 1000 simulated data > sets) as a function of ntrials. > > -- Joe, K1JT > > ------------------------------------------------------------------------------ > _______________________________________________ > wsjt-devel mailing list > wsjt-devel@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/wsjt-devel ------------------------------------------------------------------------------ _______________________________________________ wsjt-devel mailing list wsjt-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/wsjt-devel
