Subject: [Re: Frequency domain simulation]
From: Mark Zwolinski (mz@ecs.soton.ac.uk)
Date: Tue Feb 05 2002 - 06:07:30 PST
I sent this in reply to Raghuram and Parthasarathy last week. For
various tedious technical reasons it bounced from the VHDL-AMS
reflector. For the record, my reply is below - obviously if anyone else
wants to comment, feel free!
Mark
>
> Hi,
>
> Been there! Done that! Got the T shirt! Well for a SPICE-like simulator, not
> VHDL-AMS. See my comments below after each of your questions. It's probably
> easier than you think!!
>
> > hi group,
> >
> > Permit us to introduce ourselves as Parthasarathy Narasimhan and Raghuram
> > Srinivasan , students of Dr. Hal
> > Carter working in the area of Frequency Domain Simulation for VHDL-AMS.
> >
> > We seek clarification on the following subjects
> >
> > 1. In order to incorporate the Frequency solver to our existing Time
> > domain simulator - SEAMS (which
> > involves the trapezoidal integration, NR linearization and matrix solver)
> > we propose to do the following,
> >
> > a. Transalate all Simult. Stmts into a set of freq. stmts with complex
> > variable j, frequency.
> > b. Apply two-graph MNA to fill up matrix of complex nos. from a.
> > c. Solve the matrix.
> > d. write the gain responses and the phase responses of Q( quantities) and
> > f in output files.
> > e. Increment freq to the next frequency step and goto a , till reached the
> > final frequency.
> >
> > In trying to do the above we have the following queries,
> >
> > a. What is the way to convert Simultaneous statements in time/
> > frequency to frequencyonly for
> > solving them by the frequency solver?.
>
> As you note, you need a matrix of complex numbers. You need to separate real
> components from imaginary. In the time domain, the "imaginary" entries to
> the network matrix are multiplied by 1/h (or the higher order equivalent).
> In the AC (small signal frequency) domain, multiply these same entries by
> "omega" and put them in the imaginary part of the matrix. BUT you don't put
> anything on the RHS.
>
> > b. Is There a difference in the method for solving a matrix with
> > timeonly components and frequency
> > only components?.( can we use the same sparse approaches with
> > matrices in frequency, as used in time?)
>
> Well, you could create separate real and imaginary matrices and reorder them
> independently for minimizing infills. I wouldn't. Stick with the same matrix
> structure that you have for the time domain, but make the entries complex.
> Of course you'll have to define complex arithmetic operations, unless you
> use a C++ class or similar.
>
> > c. The LRM talks about a domain signal( time /frequency
> > /intialization) where is this Domain
> > signal set( for time/ frequency calculations)?
>
> I guess you set that yourselves. I've not done that bit.
>
> > d.We know that the Matrix approaches for solving come to a
> > covergence at a time point by iterating,
> > should such a pattern also be followed in frequency domain
> > solving?
> >
> > a. The iterations take place because ( of the presence of
> > non-linearities) of the
> > inclusion of a derivative or an integral operator, but
> > when we convert such a model to
> > frequency we do not have such non linearity( df(x)/dt =
> > sF(s)) thus is there a need inclusion of a derivative or an integral
> > operator, but when we convert such a model to frequency we do not have
> > such non linearity( df(x)/dt =
> > sF(s)) thus is there a need for iteration?
> >
>
> OK. You're missing an important point. You're doing linear AC analysis. You
> must have a converged DC analysis before you start the AC analysis. You have
> to calculate and keep safe all the di/dv entries and dq/dv entries (the
> latter for the imaginary part of the matrix. Of course in a DC analysis you
> won't have bothered to calculate all the dq/dv entries. Well you now have to
> do it, once). The analysis at each frequency point is now purely linear.
> Multiply each imaginary entry by omega and solve the matrix. Once. That's
> it. I noted above that the RHS is empty. The only thing that goes into the
> RHS vector is the AC stimulus. This comes from one or more VHDL-AMS source
> quantities.
>
> > e. Is there any difference between a time and the frequency domain
> > simulation if the model does not have any spectral components?.
> >
>
> Hmm. No. If you're talking about spectral components due to non-linearities,
> AC analysis (and VHDL-AMS) can't model these. See the thread about MW and
> other FD modelling on this reflector.
>
> >
> > Finally a tangential thought outside LRM,
> >
> > Could we hope to simulate models in time and frequency depending on the
> > user's choice by appropriately
> > converting them , and simulating them in time and frequency as the user
> > chooses.
> >
>
> Not sure what you mean. You should be able to write an AC model that can be
> interpreted in the time domain and vice versa. You can use the domain signal
> to switch between models. What else do you want to do?
>
> >
> > Thanks,
> >
> > Raghuram &
> > Parthasarathy
> >
>
> Mark Zwolinski
> --
> =================================================================
> Dr Mark Zwolinski Tel. (+44) (0)23 8059 3528
> Dept of Electronics & Computer Science Fax. (+44) (0)23 8059 2901
> University of Southampton Email. mz@ecs.soton.ac.uk
> Southampton SO17 1BJ, UK http://www.ecs.soton.ac.uk/~mz
-- =================================================================== Dr Mark Zwolinski Electronic System Design Group Tel. (+44) (0)23 8059 3528 Dept. of Electronics & Computer Science Fax. (+44) (0)23 8059 2901 University of Southampton Email. mz@ecs.soton.ac.uk Southampton SO17 1BJ, UK http://www.ecs.soton.ac.uk/~mz
This archive was generated by hypermail 2b28 : Tue Feb 05 2002 - 06:41:11 PST