[ibis-macro] Re: An AMI Overview
- From: <fangyi_rao@xxxxxxxxxxx>
- To: <dkirsanov@xxxxxxxxxx>, <ibis-macro@xxxxxxxxxxxxx>
- Date: Thu, 15 Oct 2009 10:57:18 -0600
Hi, Danil;
I think we are on the same page. The fundamental problem is that current AMI
spec allows two usages of Init and GetWave as stated at the end of the BIRD.
Further discussion identified that
two different styles of modeling were possible and should be supported. In the
default case, the
AMI_Init and AMI_Getwave calls represent filtering performed by sequential
stages of a device, and
the results should therefore be chained together. In the second case, the
AMI_Init and AMI_Getwave
calls each represent the overall device. For example, the AMI_Init call could
provide an LTI model
for the device while the AMI_Getwave call provides a time-varying model. In
this case, results from
the AMI_Init and AMI_Getwave calls should be treated as independent.
The first case obviously can’t support statistical simulation.
There are different approaches to make the AMI standard to support both time
domain (pattern depend) and statistical simulations. The two-model approach
suggested by Kumar is one of them. Walter and Arpad prefer to stay with one
model by dropping support to the first case in AMI since practically no model
of this usage exists.
I want to take this opportunity to suggest another solution with one model. In
AMI we should separate the interface for time domain simulation from that for
statistical by adding a third function named GetImpulseForStatistical(). It
takes impulses of victim and aggressors and returns modified impulses.
Regards,
Fangyi
From: ibis-macro-bounce@xxxxxxxxxxxxx [mailto:ibis-macro-bounce@xxxxxxxxxxxxx]
On Behalf Of Danil Kirsanov
Sent: Thursday, October 15, 2009 8:30 AM
To: ibis-macro@xxxxxxxxxxxxx
Subject: [ibis-macro] Re: An AMI Overview
Fangyi,
Could you please clarify it? I agree that it might be reasonable to put the
linear part of the model in Init and non-linear part in Getwave, so that
together they characterize the model (statistical simulator uses only the
linear part, while the pattern-dependent always uses both).
But I strongly disagree that inside one model Init and GetWave can provide
different approximations of the same algorithm (i.e. introducing
double-counting), where statistical simulator uses Init and pattern-dependent
simulator uses GetWave. I believe this behavior should be prohibited, since it
makes the flow more complicated, and we can easily achieve the same result
providing two different models (or having internal option to switch the model
between the statistical and non-linear mode).
If we have this simple rule (non-linear simulator we always uses Init and
Getwave), the behavior of the EDA does not depend on the fact whether GetWave
exists or not, and GetWaveExists flag becomes unnecessary (if the Simulator at
some point figures out there is no GetWave, it just does not use it).
Are we on the same page here?
Best,
Danil
From: fangyi_rao@xxxxxxxxxxx [mailto:fangyi_rao@xxxxxxxxxxx]
Sent: Wednesday, October 14, 2009 7:48 PM
To: kumarchi@xxxxxxxxx; ibis-macro@xxxxxxxxxxxxx; dkirsanov@xxxxxxxxxx
Subject: RE: [ibis-macro] Re: An AMI Overview
Hi, Kumar;
What if a model wants to support non-linear time domain simulation by GetWave
and statistical simulation by returning a LTI approximation in Init?
Thanks,
Fangyi
From: ibis-macro-bounce@xxxxxxxxxxxxx [mailto:ibis-macro-bounce@xxxxxxxxxxxxx]
On Behalf Of C. Kumar
Sent: Wednesday, October 14, 2009 4:23 PM
To: ibis-macro@xxxxxxxxxxxxx; dkirsanov@xxxxxxxxxx
Subject: [ibis-macro] Re: An AMI Overview
i agree..
if the model modifies the init it is the only thing it should be doing. there
should not be any getwave
--- On Wed, 10/14/09, Danil Kirsanov <dkirsanov@xxxxxxxxxx> wrote:
From: Danil Kirsanov <dkirsanov@xxxxxxxxxx>
Subject: [ibis-macro] Re: An AMI Overview
To: ibis-macro@xxxxxxxxxxxxx
Date: Wednesday, October 14, 2009, 6:59 PM
Dear colleagues,
I would like to clarify one basic principle of AMI modeling, hoping that all of
us agree with it.
I believe that the model writer should never do a double-counting: if he
modified the channel impulse response in Init() to model some effect, he should
not model this effect in Getwave(). So he cannot put the “true” model in
GetWave() and it’s linear approximation in Init(). If both types of behavior
are expected, there should be two models (or some internal flag that changes
the behavior of the model).
If this assumption is true, statistical (linear) simulator always works with
Init() function of the model, while pattern-dependent (non-linear) simulator
works with both Init() and GetWave() and I do not see any necessity for
Get_Wave_Exists flag.
Best,
Danil
Other related posts:
