Labeled gate pulse gate

[nkanazawa1989]

Summary

This PR enables to identify custom gate differentiated by the instruction label. This will be supported by #5885 in more sophisticated way. Meanwhile, a new pass LabelIdentifier (I don't like the name though) pass supports the pulse gate definition for the labeled gates.

(motivation)

This is requirements from Qiskit Experiments to support interleaved RB experiments with custom gate. Interleaved element should be defined as a gate with one of predefined names to be converted into Clifford instance (because RB cannot handle gates that don't belong to the Clifford group). To meet this requirements, we need to distinguish the interleaved gate from the standard gate with instruction labels. e.g. When we measure a CX gate that we calibrate, we want to interleave that gate. To perform IRB, we need to construct reference sequence as well, that may include another CX gate. We usually want to use backend-calibrated gate as a reference in this situation. So somehow we need to distinguish them.

(sample code)

Add labeled custom gate to instmap.

from qiskit import circuit, pulse, transpile
from qiskit.test.mock import FakeCasablanca

backend = FakeCasablanca()
instmap = backend.defaults().instruction_schedule_map

my_cx = circuit.library.CXGate(label="custom")

fake_cx_sched = pulse.ScheduleBlock()
fake_cx_sched.append(pulse.Play(pulse.Constant(1600, 0.1), pulse.ControlChannel(0)))

instmap.add(my_cx, (0, 1), fake_cx_sched)

This gate is registered with the unique name.

>>> instmap.instructions  
['cx', 'id', 'measure', 'rz', 'sx', 'u1', 'u2', 'u3', 'x', 'cx_custom']

Then, create circuit with custom gate.

qc = circuit.QuantumCircuit(2)
qc.cx(0,1)
qc.append(my_cx, [0, 1])

>>> qc.draw()
           custom 
q_0: ──■─────■────
     ┌─┴─┐ ┌─┴─┐  
q_1: ┤ X ├─┤ X ├──
     └───┘ └───┘  

The custom gate is transpiled as standard CX gate (if I set opt level 1 these gates are cancelled out).

qc_transpiled = transpile(qc, backend=backend, optimization_level=0)

The output circuit looks like the same, however, its name is replaced by cx_custom with the corresponding pulse gate extracted from the backend. This program can be actually executed on IBM backends that support pulse gate.

>>> qc_transpiled.draw()
                     custom 
      q_0 -> 0 ──■─────■────
               ┌─┴─┐ ┌─┴─┐  
      q_1 -> 1 ┤ X ├─┤ X ├──
               └───┘ └───┘  
ancilla_0 -> 2 ─────────────
                            
ancilla_1 -> 3 ─────────────
                            
ancilla_2 -> 4 ─────────────
                            
ancilla_3 -> 5 ─────────────
                            
ancilla_4 -> 6 ─────────────

>>> qc_transpiled.calibrations        
{
    'cx_custom': {
        ((0, 1), ()): ScheduleBlock(Play(Constant(duration=1600, amp=(0.1+0j)), ControlChannel(0)), name="block0", transform=AlignLeft())
    }
}

Details and comments

TODO

• unittest
• reno

[kdk]

This behavior seems specific to a particular use-case and likely not something users would want in general. e.g. This would replace all gates with a label to {name}_{label} whenever using pulse gates, even if {name}_{label} is outside the basis_gates, so I'd prefer there be an easier way for users to opt-in to this behavior when needed.

[nkanazawa1989]

That's good point. Other option would be initializing this pass with instmap and let the pass check the map if the {name}_{label} exists for the qargs (EDIT: to make sure it can be replaced by pulse gate afterward, i.e. this pass is inserted only when pulse gate pass is active). I think #6403 is good for this usecase, but seems like it will take a bit more time to merge in.

[nkanazawa1989]

Done in 1d2d540 This checks instmap before changing gate name. For example,

my_cx1 = circuit.library.CXGate(label="custom")
my_cx2 = circuit.library.CXGate(label="dummy")

instmap.add(my_cx1, (0, 1), fake_cx_sched)

qc = circuit.QuantumCircuit(2)
qc.cx(0,1)
qc.append(my_cx1, [0, 1])
qc.append(my_cx2, [0, 1])

qc_transpiled = transpile(qc, backend=backend, optimization_level=0)

>>> [inst.name for inst, _, _ in qc_transpiled.data]
['cx', 'cx_custom', 'cx']

[chriseclectic]

I don't see why you need to add a new pass that renames nodes in a DAG since this sort of thing may have issues with other passes that rely on names (ie basis gates like @kdk said).

Since you already updated the _get_instruction_string method in InstructionScheduleMap so that it's add, has and get etc methods will look for labels, can't you just update the PulseGates pass to use this information?

In PulseGates.supported and PulseGates.get_calibration it is currently passing the name string to the schedule map: https://github.com/Qiskit/qiskit-terra/blob/40deaf108b12ec7ef8c9d9ddb414a88a969f834e/qiskit/transpiler/passes/calibration/builders.py#L459 but you should just be able to just change this to pass the instruciton itself like self._inst_map.has(node_op) so it can use the updated _get_instruction_name for labelled instructions.

[nkanazawa1989]

Indeed this is what I was first trying to implement. However, the format of calibration in qobj is

Dict[name: str, Dict[Tuple[qubits: Tuple[int, ...], parameters: Tuple[float, ...]], schedule: Schedule]]

and there is no place to define the label here. Even though the InstructionScheduleMap is keyed on Instruction with a name and label, qobj doesn't distinguish two (or more) labeled instructions. In the backend, it entirely overrides calibration regardless of the label.

To use the pulse gate, we need to give unique name to the circuit instruction, then attach the calibration to the uniquely named instruction. This is why the instruction name is updated in the DAG.

[nkanazawa1989]

Close this PR. Below trick works without modification of terra.

new_inst = copy(inst)
new_inst.name = "some_new_inst_name"

qc = QuantumCircuit(inst.num_qubits, inst.num_clbits)
qc.append(inst, range(inst.num_qubits), range(inst.num_clbits))

new_inst._definition = qc