Improve circuit pre-processing for `Cirq` converter with existing gate decomposers

[ryanhill1]

Before passing an input cirq circuit to the QIR converter, we first decompose the circuit to ensure that it only uses gates / operations that are supported by QIR, see PYQIR_OP_MAP. Right now, to accomplish this, we use a preprocess_circuit function which loops through each operation, and uses a try / except to determine if the operation is supported, and if not, applies a naive cirq.decompose_once to the circuit.

qbraid-qir/qbraid_qir/cirq/passes.py

Lines 34 to 43 in 3ef504f

	try:
	# Try converting to PyQIR. If successful, keep the operation.
	_ = map_cirq_op_to_pyqir_callable(operation)
	return [operation]
	except CirqConversionError:
	pass
	new_ops = cirq.decompose_once(operation, flatten=True, default=[operation])
	if len(new_ops) == 1 and new_ops[0] == operation:
	raise CirqConversionError("Couldn't convert circuit to QIR gate set.")
	return list(itertools.chain.from_iterable(map(_decompose_gate_op, new_ops)))

For the majority of use-cases, this works. However, it is:

1. Bad style
2. Inefficient

Instead, we would like to use one of the built-in Cirq transforms such as cirq.optimize_for_target_gateset to automatically ensure that the input circuit conforms to the supported QIR operations. In doing so, hopefully we can eliminate the try except blocks in passes.py, and re-implement preprocess_circuit without the clunky helpers.

[arulandu]

@ryanhill1 Is this still open? I'd love to work on this!

[ryanhill1]

@arulandu It is, go for it!

[arulandu]

I'm thinking of creating a Gateset by extending https://quantumai.google/reference/python/cirq/TwoQubitCompilationTargetGateset. However, I'm unsure how to implement self._decompose_two_qubit_operation. Could we just use https://quantumai.google/reference/python/cirq/CZTargetGateset?

[ryanhill1]

I'll admit, I'm not super familiar with how cirq.CompilationTargetGateset work. It seems like you just start by passing in a list of cirq.Gate, so to start, this would be the set gates supported by map_cirq_op_to_pyqir_callable(), I think? And then, perhaps there's code from qbraid_qir/qasm3/maps.py and/or qbraid_qir/qasm3/linalg.py that we can move to the top-level and generalize to expand that list or assist with some of the decompositions? What gates are included in the cirq.CZTargetGateset? If it maps to the pyqir gateset that we are targeting closely, I guess that could be a reasonable starting point.

@TheGupta2012 @skushnir123 perhaps you guys could take a look here and weigh in?

Or if @vtomole has any insights, as someone who is much closer to the Cirq stack, that could be super helpful as well.

[arulandu]

cirq.CZTargetGateset is cirq.CZPowGate, PhasedXZGate, MeasurementGate, GlobalPhaseGate. Yes, it should indeed take the list of gates, but you need to implement the abstract method for decomposing gates. TwoQubitCompilationTargetGateset does some of the work and only requires you to implement a 2q unitary decomposition. Should we be implementing a Clifford-T decomposition like here and then replace iSWAP with SWAP + Xs?

[vtomole]

What gateset do ya'll want to decompose to? The list in PYQIR_OP_MAP is kinda big. You can just decompose to Clifford + T if the the input gate is not in that list.

I'm unsure how to implement self._decompose_two_qubit_operation

Why not copy CZTargetGateset's implementation?

[arulandu]

I would copy their implementation and I could decompose PhasedXZ into {Rx,Rz}, but I'm unsure what to do about the
CZPowGate and GlobalPhaseGate

[vtomole]

CZ is in PYQIR_OP_MAP. Why not call two_qubit_matrix_to_cz_operations like CZTargetGateset does?

[ryanhill1]

You can leave out GlobalPhaseGate as there's not a way to represent that in pyqir at the moment (to my knowledge).

And yes, like Victory said, the "CZ" gate is already accounted for in our current implementation, so if the CZPowGate exponent matches to a regular Pauli Z, then we can just map that directly to pyqir._native.cz. Otherwise, you can look at it like ControlledGate(ZPowGate) and follow the different ZPowGate exponent cases that are given in opsets.py:

qbraid-qir/qbraid_qir/cirq/opsets.py

Lines 112 to 118 in 4f009c3

	elif isinstance(gate, (cirq.ops.XPowGate, cirq.ops.YPowGate, cirq.ops.ZPowGate)):
	if gate.exponent == 1 or (
	isinstance(gate, cirq.ZPowGate) and gate.exponent in [0.25, -0.25, 0.5, -0.5]
	):
	op_name = str(gate) # X, Y, Z, S, T, S**-1, T**-1
	else:
	op_name = f"R{gate.__class__.__name__[0].lower()}" # Rotations