Binding a model to a design

One classmethod on the design says which model prices it:

class VecMult(FreeRunMod):

    @classmethod
    def get_rm(cls, platform):
        ...

Return None — or do not define it at all — to take the default lookup against the platform’s measurement store. That is the right answer whenever you can afford to measure every configuration you will ask about, so most designs write nothing here.

A model that derives its counters also declares resource_structure(); that is a VitisResourceModel concept and lives with it.

get_rm(platform) — which model, on this platform

@classmethod
def get_rm(cls, platform):
    part = getattr(platform, "part", None) or PART
    require_same_device(part, PART, what="VecMult's resource model")
    store = ModuleStore(getattr(platform, "dir", None) or COMMITTED_CALIB)
    return VitisResourceModel(
        name="vec_mult", part=part, platform=platform,
        cls_name="VecMult", comp_class=cls, store=store,
    ).load_or_fit()

Why a classmethod

Because a model must not close over an instance, and having no self makes that impossible rather than merely discouraged.

The model is handed the component to predict for. The same object has to price every point of a corpus during fit and every sibling during compose — bind it to one instance and every row of the fit becomes identical, silently.

Everything configuration-specific still reaches the model, just later: through resource_structure() on whatever component it is asked about, at predict time.

The key is (class, platform) — not the parameters

The base caches what get_rm returns:

bound to an instance      ✗   breaks fit() and compose()
a class variable          ✗   coefficients depend on the platform
keyed (class, platform)   ✓   one object, cached, prices every configuration

Parameters are absent from the key, and that is a direct consequence of the model being instance-agnostic. One VitisResourceModel for VecMult prices dwid=64, vlen=4096 and dwid=256, vlen=1024 equally well. Had the structure been bound, the key would have needed every parameter and the cache would be one entry per design point.

Refuse the wrong platform

get_rm is where a platform this class cannot be modelled on gets rejected. Returning a model that silently applies another technology’s geometry is the worse failure — see guarding the part.

What the base does with it

top.add_rm(platform)     # once, on the top — post-order over the whole hierarchy

For each module: resolve get_rm (cached), install it, or fall back to the store lookup. A module with no model contributes zero and reports UNCALIBRATED — never silently skipped, because a missing contribution makes a design read as cheaper than it is.

Next

• Predicting — turning installed models into an estimate.
• Fitting — where the coefficients in load_or_fit come from.