nonuniform cost of physical swaps in sabre

[ajavadia]

Summary

Modifying the distance matrix between physical qubits in SabreLayout to account for the cost of swapping each pair of qubits. This cost is inferred from Target by looking at the available set of gates on that pair of qubits, their errors, and how many of them are needed to do a SWAP. The scoring of each swap is thus affected by the information in Target. Requires #9263.

Details and comments

Example:

import numpy as np
from qiskit.circuit.library import HGate, CXGate, RZZGate, SwapGate
from qiskit.transpiler import Target, InstructionProperties
from qiskit import QuantumCircuit, transpile
from qiskit.transpiler import Target

# build circuit
c = QuantumCircuit(3)
c.cx(0, 1)
c.cx(0, 2)
c.cx(2, 1)
c.cx(2, 0)
c.cx(2, 1)
c.cx(0, 1)
c.cx(0, 2)
print(c)

# build target
target = Target()
h_properties = {
    (0, ): None, (1, ): None, (2, ): None
}
swap_properties = {
    (0, 1): InstructionProperties(error=.1),
    (1, 2): InstructionProperties(error=.2)
}
cx_properties = {
    (0, 1): InstructionProperties(error=.1),
    (1, 2): InstructionProperties(error=.2)
}
rzz30_properties = {
    (0, 1): InstructionProperties(error=.03),
    (1, 2): InstructionProperties(error=.07)
}
target.add_instruction(HGate(), h_properties)
# any of the following two-qubit gates can be present on any pair
target.add_instruction(CXGate(), cx_properties, name='cx')
target.add_instruction(SwapGate(), swap_properties, name='swap')
target.add_instruction(RZZGate(np.pi/6), rzz30_properties, name='rzz30')

new_c = transpile(c, target=target, layout_method='sabre', seed_transpiler=1, optimization_level=0)

error SWAP(0, 1) = error SWAP(1, 2)

error SWAP(0, 1) > error SWAP(1, 2)

error SWAP(0, 1) < error SWAP(1, 2)

[qiskit-bot]

Thank you for opening a new pull request.

Before your PR can be merged it will first need to pass continuous integration tests and be reviewed. Sometimes the review process can be slow, so please be patient.

While you're waiting, please feel free to review other open PRs. While only a subset of people are authorized to approve pull requests for merging, everyone is encouraged to review open pull requests. Doing reviews helps reduce the burden on the core team and helps make the project's code better for everyone.

One or more of the the following people are requested to review this:

• @Eric-Arellano
• @Qiskit/terra-core
• @enavarro51
• @kevinhartman
• @mtreinish

[mtreinish]

Do you think that it makes sense to wrap this in a try/except if the target basis isn't known to TwoQubitWeylDecomposition and fall back to the default distance matrix? I'm just wondering what happens if the decomposer doesn't handle the 2q gates in the basis?

[ajavadia]

yeah you are right i still need to write test cases for the different cases we may encounter

[mtreinish]

We might want to put this in an else condition because this will probably compute the distance matrix (it is cached so if it was used before it won't) which can be expensive depending on the size of the graph. We probably don't want to do build two distance matrices if we're only going to use one.

[mtreinish]

The other question I just thought of is how does this interact with VF2PostLayout? Since we might be selecting different physical qubits after we do routing the error rates we're using for swaps might not hold true for the new layout used. It might not matter in the practical sense because vf2postlayout will only pick a matching layout that has lower aggregate error. But I'm just wondering of the impact of the heuristic to be noise aware like this in the context where we may select an alternative layout after we run sabre routing.

[jakelishman]

I'll close this as stale - we'd have to completely rewrite the PR if it were going to merge.