How can I save the converged values at the nodes after the last iteration of the time step in verilog-A?

I don't really understand from your example what you're expecting it to do (it seems a rather strange requirement from your initial description). Why would you need to do this?

Andrew

I don't really understand from your example what you're expecting it to do (it seems a rather strange requirement from your initial description). Why would you need to do this?

Andrew

Sorry for the confusion.

I have a set of equations in the form of:
x1 = f(h_prev)
h = g(x1)
where h_prev is h (the converged value) at the previous time step and initially assumed to be 0, g and f are functions.
so I need to save the converged value of h at every time step and assign it to h_prev at the begin of the next time step (excluding the initial one)

[Update]

I thought about writing the values to a file and then read them but I found a fatal because I'm trying to change the mode of the file. Is there a way to read and write from the same file?

That seems very odd, because if you have variable time steps (which is usual with spectre), your function is going to have this strange dependency on the varying timestep. I could understand it if you were using @(timer...) to schedule a regular even and update the variables that way (i.e. like modelling some sampled data system).

If you really want this to behave this way with a variable time step you could use something like:

if ($abstime != lastTime) begin
  h_prev=h;
  lastTime = $abstime;
end

... use h_prev in calculations and compute new h

I've not tested this (mainly because I can't think of an application of it), but I think it should work 

Andrew

I agree that the idea is very odd.

So basically I have got this ODE from a well-trained ML model that predicts a circuit behavior:

dh_dt(h(t), t) = w2 * tanh( w1 * h(t) + b1) + b2 

where: {w1, w2, b1, b2} is the parameter set of a neural network and it's known
            h(t) is the circuit state at t and dh_dt represents the dynamics (the rate of change)

I want to solve an IVP at $abstime when h0 (the initial state) is known so I'm using:

V(h_final) <+ idt( w2 * tanh( w1 * V(h) + b1) + b2, h0, lastTime); 
then, I have an equation where V(ht) is a function of V(h_final)

where: h, h_final are electrical nodes
           h0 is updated by the last value of V(ht)

but when I debug, I'm always getting V(h_final) = 0.0, Do you think the problem with the integral? 

Debugging this without seeing the full model is rather difficult (and to be honest I'm unlikely to have time to do that in the coming days before vacation). So the best thing to do with this is to contact customer support where you can share the whole model and a related testbench.

Andrew

Thanks, Andrew for your time and advice.

One last question if possible:

I need to do integration on a certain interval, e.g. t1 -> t2, then do some calculation on the output of integration and use the result as an initial condition for subsequent integration e.g. t2 -> t3 and so on

I tried to do that as follows:

//initialized ic and dh_dt
integer ic = 0;
real dh_dt = 1.591112;
real dh, h_, h;
real t_scale = 0.2 * 1e6;
real t, t_prev;

analog begin
t = $abstime * t_scale;
h_ = idt(t_scale * dh_dt, ic) + dh;
if (t != t_prev) begin
// .. some calc. on h_ to get h and new dh_dt
dh = h - h_; 
t_prev = t;
end

But that wasn't efficient, is there a direct way to do so?