Re: gEDA-user: Periodic steady state in NGSpice

[al davis <ad106@xxxxxxxxxxxxxxxx>]

On Tuesday 19 September 2006 09:02, Dan McMahill wrote:
> I'll ask a naive question here.  I thought that the way
> shooting methods worked were after the initial transient you
> took a snapshot of the state, and ran for what you believe to
> be exactly 1 period and then compared the state of the
> circuit again and if it didn't match, you do sort of newtons
> method calculation to pick a new starting state.

I said "approximately".  The idea of changing initial conditions 
is a way of avoiding the transient run to get past the start-up 
transient.  Hopefully, this would make it faster.

> Or maybe you're just describing a different way of deciding
> when you've run enough cycles in a transient sim to find the
> periodic steady state solution.

Yes ... but the advanced algorithms try to avoid that transient 
sim to get past the start-up, into steady state.  I was 
describing a quick and dirty way to get there.

> I don't think I quite understand about the tweaking the
> frequency and running the fourier analysis again bit.

You need to run it for "exactly 1 period".  The gnucap fourier 
command first runs a transient analysis with the stepping 
chosen to be optimum for the frequencies you 
selected.  "Tweaking the frequency" means that after running it 
once, you adjust the frequency slightly with the goal of really 
getting "exactly 1 period".

> I'll add a caution that if you're not comparing the state of
> the entire circuit you have to be very careful to observe a
> signal which should have the fundamental period of the
> circuit as a whole and also pick one with the longest time
> constants.  The first is important or you might think you
> have a periodic solution but suppose there is some block
> which creates subharmonics (a toggle flip-flop somewhere for
> example), you may have found a half-period solution instead
> of the whole period.  Also suppose you have some bypassed
> bias line or some slow control loop. That line may not even
> have much of the periodic signal on it but it may have a long
> initial transient that takes a while to die out.
>
> > In ngspice, you will need a real Fourier transform, and the
> > ability to do repeat runs, with changes between them in
> > such as way as to not start over every time.
>
> This is a very nice feature of gnucap.
>
> > You will also need to
> > find a way to synchronize the data points with the Fourier
> > points, otherwise you will see what appears to be a high
> > noise floor, even if your frequencies are perfectly
> > matched.  Also, I question whether the step size control is
> > capable of getting acceptable accuracy.  You may need to
> > completely replace the time step control code.
>
> without having looked, does ngspice have a "strobeperiod"
> parameter like spectre?  That basically says "in addition to
> your usual time step algorithm, output data points every
> strobeperiod seconds being sure to actually solve the circuit
> here"

I don't think so.  As I recall, it gets the Fourier points by 
interpolating on a previous transient analysis.  This is a big 
source of error.  In gnucap, the time step you specify is 
interpreted the same as the spectre strobeperiod.  Based on a 
conversation I had recently with some of the spectre people, I 
think gnucap does a better job at taking that into account when 
it picks the internal time steps.  Gnucap tries hard to keep 
the time step constant, while still accomodating a need for 
changing the time step.

As an example ..  suppose you ask for stepping every 
microsecond, and the error control says you need .4 
microsecond.  Gnucap will give you .3333333 consistently, so 
every third step matches the requested ones.  I believe, but 
don't know for sure, that spectre gives 
you .4, .4, .2, .4, .4, .2 etc.  I believe that spice 3f5 (and 
therefore ng-spice) would give you .4 repeated,missing the 
strobe points.  Actually, I believe ng-spice would miscalculate 
the needed stepping and give you 1 us repeated, with none of 
the points matching the strobe, so it  would interpolate every 
time. 

On the other hand, suppose error control asks for 4 us.  Gnucap 
and spectre both give you the 1 us you asked for.  Spice gives 
you 1 us stepping, but makes no attempt to synchronize to the 
strobe.



_______________________________________________
geda-user mailing list
geda-user@xxxxxxxxxxxxxx
http://www.seul.org/cgi-bin/mailman/listinfo/geda-user