2 Qubit ControlledOperations should be supported in convert_to_sycamore_gates

[tanujkhattar]

Description of the Issue

We check isinstance(op, GateOperation) while verifying that the given operation can be decomposed using analytical / tabulation methods. This leads to the problem that we don't recognize two qubit controlled operations as known gates and hence don't support them.

Cirq/cirq-google/cirq_google/optimizers/convert_to_sycamore_gates.py

Line 111 in 12530f1

elif len(op.qubits) == 2 and isinstance(op, ops.GateOperation):

For example:

num_qubits = 3
qubits = cirq.GridQubit.rect(1, num_qubits, 4, 4)
circuit = cirq.Circuit(cirq.QuantumFourierTransformGate(num_qubits)(*qubits))
cirq.Circuit(cirq.google.ConvertToSycamoreGates().convert(circuit), device = cirq.google.Sycamore)

The above code fails with the following error:

TypeError: Don't know how to work with cirq.S(cirq.GridQubit(4, 4)).controlled_by(cirq.GridQubit(4, 5)). It isn't a native xmon operation, a 1 or 2 qubit gate with a known unitary, or composite.

The reason it fails is because ControlledOperations don't satisfy the isinstance(gate, GateOperation) constraint.

op = cirq.S(cirq.GridQubit(4, 4)).controlled_by(cirq.GridQubit(4, 5))
print(type(op), isinstance(op, cirq.GateOperation))

gives

<class 'cirq.ops.controlled_operation.ControlledOperation'> False

However, the gate is actually just a two qubit controlled gate and should be decomposable into the native gateset via known_two_q_operations_to_sycamore_operations.

op = cirq.S(cirq.GridQubit(4, 4)).controlled_by(cirq.GridQubit(4, 5))
print([*known_two_q_operations_to_sycamore_operations(op.qubits[0], op.qubits[1], op, None)])

gives the following decomposition

[cirq.PhasedXZGate(axis_phase_exponent=-1.1102230246251565e-16, x_exponent=0.7239292897100449, z_exponent=0.08333333333333348).on(cirq.GridQubit(4, 5)), cirq.PhasedXZGate(axis_phase_exponent=0.5416666666666665, x_exponent=0.0, z_exponent=0.916666666666667).on(cirq.GridQubit(4, 4)), cirq_google.SYC.on(cirq.GridQubit(4, 5), cirq.GridQubit(4, 4)), cirq.Rx(rads=0.5922772837339743).on(cirq.GridQubit(4, 4)), cirq_google.SYC.on(cirq.GridQubit(4, 5), cirq.GridQubit(4, 4)), cirq.PhasedXZGate(axis_phase_exponent=-0.08333333333333315, x_exponent=0.7239292897100451, z_exponent=-0.916666666666667).on(cirq.GridQubit(4, 5)), cirq.PhasedXZGate(axis_phase_exponent=0.22741638234956674, x_exponent=2.220446049250313e-16, z_exponent=0.25).on(cirq.GridQubit(4, 4)), cirq.Rz(rads=0.7853981633974483).on(cirq.GridQubit(4, 5)), cirq.Rz(rads=0.7853981633974483).on(cirq.GridQubit(4, 4))]

Proposed Solution

Cirq/cirq-google/cirq_google/optimizers/convert_to_sycamore_gates.py

Line 111 in 12530f1

elif len(op.qubits) == 2 and isinstance(op, ops.GateOperation):

should be modified to also recognize two qubit ControlledOperations.

Cirq version
0.11.0.dev

Part of #3242

[smitsanghavi]

I'm assuming the GateOperation check is supposed to filter out Operations that don't have gates.

We have had these discussions about Operation hierarchy and how the type checks are used as proxies for whether there is an underlying Gate or not. I think it's useful to have a has_gate protocol that actually tries to get the underlying gate and performs not-None, etc checks. This would be more pythonic than type checking IMO.

[maffoo]

op.gate is not None seems about as simple as cirq.has_gate(op), and I think that's basically all the has_gate method would do anyway. Certainly the isinstance(op, cirq.GateOperation) checks are too restrictive in this case.

[smitsanghavi]

The protocol will also be doing a getattr check for gate first. op.gate is not None assumes that op is some Operation type which might not be the case for all instances of these checks in Cirq.

[tanujkhattar]

Upd: Similarly, _is_native_xmon_op checks for isinstance(op, cirq.GateOperation) and hence does not recognize gates like CS(0, 1) to be native gates which leads to much worse (decomposed) target circuit lengths. xref #4172

[vtomole]

Discussed in Cirq sync: Recent work on the circuit transformer APIs should fix this.