Feature request: re-usable components

[jhenin]

In a first stage, these may be cvc objects that are thin wrappers around other cvcs and only implement getting the values, not applying forces.

[HanatoK]

Just curious about this feature. Does this mean we can do something like compute RMSD of dihedrals or other CVs without using custom functions or tcl scripts?

[jhenin]

For that I think you'll still need custom functions or scripts. This feature is useful when several variables are needed that are different combinations of the same components. Then they'll only need to be computed once. In a first "read-only" stage, it will allow doing analysis on individual components while using them in a separate colvar for biasing.

[giacomofiorin]

@HanatoK To compute a RMSD of a set of dihedral there is a dihedralPC component, which assumes input in the format used by CARMA: https://utopia.duth.gr/glykos/carma.html and uses dihedral components. You can probably make a similar class with very small effort, should the scripted or custom function be too slow.

[giacomofiorin]

Looks like this needs #254 done before moving ahead. Assign back to me if needed.

[jhenin]

Hi @giacomofiorin This feature is getting more useful with additions by @HanatoK (PR coming soon). Another possible pathway is to make a colvar depend on another colvar, but this seems more lightweight in terms of architecture changes.

[giacomofiorin]

Good to know, I look forward to the PR.

I also agree that it's better to keep formalize the dependencies between components, since we have already a few ad-hoc cases and it is the components that take up the bulk of the computation.

[HanatoK]

Hi @giacomofiorin,
I don't know too much about colvardeps. Is that possible to treat the dependencies between components just the same as colvarbias and the components? If that can be implemented, then we can refer a CVC in different definition of colvar {} blocks to avoid redundant computations.

[giacomofiorin]

Yes, what you ask is possible in theory but not implemented yet.

[jhenin]

The current dependency system would allow a CVC to have several parent colvars, just the same way a colvar can have several parent biases.

[giacomofiorin]

The current dependency system would allow a CVC to have several parent colvars, just the same way a colvar can have several parent biases.

But the loop over CVCs in the SMP version does not take that into account, and could easily end up invoking the same CVCs on multiple threads at the same time.

It's not a big change to work around that, but it needs to be done.