[NPU] Use zero tensor to get correct data

[pereanub]

Details:

• Can not deallocate and re-allocate a newer memory for tensor if update mutable command list is not supported. Newer memory should be updated in the graph and this can be done only using the updating command list feature
• In order to check if the memory was or wasn't re-allocated we are checking the unique ID provided by the driver when memory is created

Tickets:

• E#134453

[razvanapetroaie]

Halfway through, nothing major so far.

[razvanapetroaie]

@MirceaDan99 Please let us know if you have anything before merging this.

[razvanapetroaie]

Since these three ifs share the same body, you may consider writing them in a more compact manner. That is: if (condition1 || condition2 || condition3) then common_body.

I've also noticed you're using std::dynamic_pointer_cast<ZeroRemoteTensor> quite often for determining whether the tensor is remote or not. You may consider creating a static method inside ZeroRemoteTensor, perhaps called is_remote_tensor, which does this check.

[pereanub]

Not sure about the second part of the comment. We still need to do the cast, correct? Since it is an ITensor as default. Can I call methods from the derivate class in this case?

[razvanapetroaie]

Was thinking you could do something like:

class ZeroRemoteTensor {
public:
static bool is_remote_tensor(const std::shared_ptr<ov::ITensor>& tensor) {
    return std::dynamic_pointer_cast<ZeroRemoteTensor>(tensor) != nullptr;
}
};

And call it like: ZeroRemoteTensor::is_remote_tensor(tensor). So I guess in this case you would just place the helper function inside the ZeroRemoteTensor "namespace" practically. Not sure if this is the best approach or the best place for the helper function, the suggestion is mainly about defining a helper for this wherever you think it belongs best.

[razvanapetroaie]

We sure we should skip state tensors? Just asking if you've considered this case hard enough, I haven't checked this yet. The tensor can be retrieved via SyncInferRequest::query_state -> get_state.

[pereanub]

Not sure yet how this is supposed to work with state tensors.