Add Moment.from_ops to more efficiently construct moments

[maffoo]

This adds a new classmethod cirq.Moment.from_ops for constructing a moment from operations without calling flatten_to_ops. This can save a significant amount of time in workloads that construct a lot of circuits. For example, in one internal benchmark we spend almost 3% of the total runtime in the cirq.Moment construct (20s out of 676s total).

This also adds functions cirq.m and cirq.c as shorthands for cirq.Moment.from_ops and cirq.FrozenCircuit.from_moments, repsectively. These are similar to the cirq.q helper for constructing qubit types (#5181) while trying to guide users to use cirq in the most performant way: avoiding the OP_TREE overhead associated with the default moment and circuit constructors, and preferring FrozenCircuit which is immutable and so can cache computed properties of the circuit. These single-letter aliases may be more controversial, but I thought I'd at least propose them. They could also be split out from this PR because I think cirq.Moment.from_ops is useful even without the helper functions.

[mpharrigan]

+1 from ops. I've hacked in my own constructor in the past for performance reasons; would be nice to have this.

The single letter things seem not very readable. Even if I grant that you can tell cirq.m(...) is creating a moment, I wouldn't guess that it is the more strict ops-only version. I would probably guess the opposite: shortcut is for general purpose and spelling things out gives you general behavior.

[dstrain115]

Agree with Matt. I think the from_ops looks fine to me, but I don't think we should include the one-letter short-hand. It's too ambiguous and having the short-hand bypass validation seems risky.

[tanujkhattar]

+1 to everything said above by Doug and Matt.

If you want the shorthand to bypass validation, I'm fine with cirq.m(*ops, flatten=False); though the character savings are marginal at this point.

[mpharrigan]

cc = CirqVerbatimBuilder()

circuit = cc.c([
  cc.m([cirq.H]),
  cc.m(ops),
])

c = CirqVerbatimBuilderWithOperatorOverrides()

# use `__call__` and `__item__`
circuit = c([
  c[cirq.H, cirq.Z],
  c[ops]
])

[maffoo]

Thanks for the comments, everyone. I've stripped out the cirq.c and cirq.m helpers and just kept cirq.Moment.from_ops. PTAL.

[maffoo]

Even if I grant that you can tell cirq.m(...) is creating a moment, I wouldn't guess that it is the more strict ops-only version. I would probably guess the opposite: shortcut is for general purpose and spelling things out gives you general behavior.

I agree with this in principle. Unfortunately we tried to cram a lot much flexibility into the constructors of basic data structures like Moment and Circuit, rather than putting the flexibility into helpers, so we pay a lot of extra cost whenever we construct these objects. I think it would be nice if we can build a better set of "default" APIs that are more strictly typed and avoid a lot of validation overhead, but maybe we can continue to explore this outside of cirq core like with the builder you showed.

[tanujkhattar]

Should we update the serialization to also include this flag? If not, deserialization would still be inefficient even if the serialized list of operations is flat.

[maffoo]

Good idea. Done.

[mpharrigan]

Unfortunately we tried to cram a lot much flexibility into the constructors of basic data structures

yeah, I likewise agree that we messed up our constructors :/

[dabacon]

doc that this doesn't do any flattening

[maffoo]

Added docstring.