Hi Steve,
Thanks for catching and fixing my error in the scaling factor for
FFT-based convolutions. I've done some further cleanup of the code in
subroutine subtract65(), which I think is now in good shape. As you
suspected, I now get even more decodes from the second pass:
Program Decodes Time
--------------------------------------------
WSJT-X v1.6.0 v5636 (kvasd) 2270 230 s
jt65 (1 pass, sfrsd2) 2296 767
jt65 (2 passes, sfrsd2) 2819 2144 *incorrect scaling*
jt65 (2 passes, sfrsd2) 2906 2748 corrected
> It might save even more time if we calculated the averaged complex
> amplitude at, say, 1/100’th or even 1/1000’th as many points enabling
> us to use a much smaller fft for the return to the time-domain. We
> could interpolate or even use a piece-wise constant approximation
> to the complex amplitude when calculating the reconstructed signal.
> Do you think that this is worth pursuing?
I'd say probably not. At present timer.out shows only 3% of execution
time is spent doing those long FFTs.
I'm thinking now about ways to give a user some flexibility on how best
to use some of the new capabilities. It seems that the best choice of
some parameters and techniques (ntrials, 2-pass, experience-based
decoding) will be different in distinct situations such as MF, HF, EME, etc.
-- Joe, K1JT
------------------------------------------------------------------------------
_______________________________________________
wsjt-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/wsjt-devel
