טרם תורגם

דף זה טרם תורגם. התוכן מוצג באנגלית.

Quantum circuit optimization

הערה

Toshinari Itoko (21 June 2024)

Download the pdf of the original lecture. Note that some code snippets might become deprecated since these are static images.

Approximate QPU time to run this experiment is 15 s.

(Note: Some cells of part 2 are copied from the notebook "Qiskit Deep dive", written by Matthew Treinish (Qiskit maintainer))

# !pip install 'qiskit[visualization]'
# !pip install qiskit_ibm_runtime qiskit_aer
# !pip install jupyter
# !pip install matplotlib pylatexenc pydot pillow

import qiskit

qiskit.__version__

'2.0.2'

import qiskit_ibm_runtime

qiskit_ibm_runtime.__version__

'0.40.1'

import qiskit_aer

qiskit_aer.__version__

'0.17.1'

1. Introduction

This lesson will address several aspects of circuit optimization in quantum computing. Specifically, we will see the value of circuit optimization by using optimization settings built into Qiskit. Then we will go a bit deeper and see what you can do as an expert in your particular application area to build circuits in a smart way. Finally, we will take a close look at what goes on during transpilation that helps us optimize our circuits.

2. Circuit optimization matters

We first compare the results of running 5-qubit GHZ state ($\frac{1}{\sqrt{2}} \left( |00000\rangle + |11111\rangle \right)$) preparation circuits with and without optimization.

from qiskit.circuit import QuantumCircuit
from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager
from qiskit.primitives import BackendSamplerV2 as Sampler

from qiskit_ibm_runtime.fake_provider import FakeBrisbane

backend = FakeBrisbane()

We first use a GHZ circuit naively synthesized as follows.

num_qubits = 5

ghz_circ = QuantumCircuit(num_qubits)
ghz_circ.h(0)
[ghz_circ.cx(0, i) for i in range(1, num_qubits)]
ghz_circ.measure_all()
ghz_circ.draw("mpl")

2.1 Optimization level

There are 4 available optimization_levels from 0-3. The higher the optimization level the more computational effort is spent to optimize the circuit. Level 0 performs no optimization and just does the minimal amount of work to make the circuit runnable on the selected backend. Level 3 spends the most amount if effort (and typically runtime) to try to optimize the circuit. Level 1 is the default optimization level. We transpile the circuit without optimization (optimization_level=0) and with optimization (optimization_level=2). We see a big difference in the circuit length of transpiled circuits.

pm0 = generate_preset_pass_manager(
    optimization_level=0, backend=backend, seed_transpiler=777
)
pm2 = generate_preset_pass_manager(
    optimization_level=2, backend=backend, seed_transpiler=777
)
circ0 = pm0.run(ghz_circ)
circ2 = pm2.run(ghz_circ)
print("optimization_level=0:")
display(circ0.draw("mpl", idle_wires=False, fold=-1))
print("optimization_level=2:")
display(circ2.draw("mpl", idle_wires=False, fold=-1))

optimization_level=0:

optimization_level=2:

2.2 Exercise

Try optimization_level=1 as well and compare the resulting circuit with the above two. Try it by modifying the code above.

Solution:

pm1 = generate_preset_pass_manager(
    optimization_level=1, backend=backend, seed_transpiler=777
)
circ1 = pm1.run(ghz_circ)
print("optimization_level=1:")
display(circ1.draw("mpl", idle_wires=False, fold=-1))

optimization_level=1:

Run on a fake backend (noisy simulation). See Appendix 1 for how to run on a real backend.

# run the circuits on the fake backend (noisy simulator)
sampler = Sampler(backend=backend)
job = sampler.run([circ0, circ2], shots=10000)
print(f"Job ID: {job.job_id()}")

Job ID: 93a4ac70-e3ea-44ad-aea9-5045840c9076

# get results
result = job.result()
unoptimized_result = result[0].data.meas.get_counts()
optimized_result = result[1].data.meas.get_counts()

from qiskit.visualization import plot_histogram

# plot
sim_result = {"0" * 5: 0.5, "1" * 5: 0.5}
plot_histogram(
    [result for result in [sim_result, unoptimized_result, optimized_result]],
    bar_labels=False,
    legend=[
        "ideal",
        "no optimization",
        "with optimization",
    ],
)

3. Circuit synthesis matters

We next compare the results of running two differently synthesized 5-qubit GHZ state ($\frac{1}{\sqrt{2}} \left( |00000\rangle + |11111\rangle \right)$) preparation circuits.

# Original GHZ circuit (naive synthesis)
ghz_circ.draw("mpl")

# A cleverly-synthesized GHZ circuit
ghz_circ2 = QuantumCircuit(5)
ghz_circ2.h(2)
ghz_circ2.cx(2, 1)
ghz_circ2.cx(2, 3)
ghz_circ2.cx(1, 0)
ghz_circ2.cx(3, 4)
ghz_circ2.measure_all()
ghz_circ2.draw("mpl")

# transpile both with the same optimization level 2
circ_org = pm2.run(ghz_circ)
circ_new = pm2.run(ghz_circ2)
print("original synthesis:")
display(circ_org.draw("mpl", idle_wires=False, fold=-1))
print("new synthesis:")
display(circ_new.draw("mpl", idle_wires=False, fold=-1))

original synthesis:

new synthesis:

The new synthesis produces a shallower circuit. Why?

This is because the new circuit can be laid out on linearly connected qubits, so on IBM® Brisbane's heavy-hexagon coupling graph as well, while the original circuit requires star-shaped connectivity (a degree-4 node) and hence cannot be laid out on the heavy-hex coupling graph, which has nodes at most degree 3. As a result, the original circuit requires qubit routing that adds SWAP gates, increasing the gate count.

What we have done in the new circuit can be seen as a manual "coupling constraint-aware" circuit synthesis. In other words: manually solving circuit synthesis and circuit mapping at the same time.

# run the circuits
sampler = Sampler(backend=backend)
job = sampler.run([circ_org, circ_new], shots=10000)
print(f"Job ID: {job.job_id()}")

Job ID: 19d635b0-4d8b-44c2-a76e-49e4b9078b1b

# get results
result = job.result()
synthesis_org_result = result[0].data.meas.get_counts()
synthesis_new_result = result[1].data.meas.get_counts()

# plot
sim_result = {"0" * 5: 0.5, "1" * 5: 0.5}
plot_histogram(
    [
        result
        for result in [
            sim_result,
            unoptimized_result,
            synthesis_org_result,
            synthesis_new_result,
        ]
    ],
    bar_labels=False,
    legend=[
        "ideal",
        "no optimization",
        "synthesis_org",
        "synthesis_new",
    ],
)

In general, circuit synthesis depends on application and it's too difficult for a software to cover all possible applications. Qiskit transpiler happens to have no functions of synthesizing GHZ state preparation circuit. In such a case, manual circuit synthesis as shown above is worth considering. In this section, we look into the details of how Qiskit transpiler works using the following toy example circuit.

# Build a toy example circuit
from math import pi
import itertools
from qiskit.circuit import QuantumCircuit
from qiskit.circuit.library import excitation_preserving

circuit = QuantumCircuit(4, name="Example circuit")
circuit.append(excitation_preserving(4, reps=1, flatten=True), range(4))
circuit.measure_all()

value_cycle = itertools.cycle([0, pi / 4, pi / 2, 3 * pi / 4, pi, 2 * pi])
circuit.assign_parameters(
    [x[1] for x in zip(range(len(circuit.parameters)), value_cycle)], inplace=True
)
circuit.draw("mpl")

3.1 Draw the entire Qiskit transpilation flow

We look into the transpiler passes (tasks) for optimization_level=1.

from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager

# There is no need to read this entire image, but this outputs all the steps in the transpile() call
# for optimization level 1
pm = generate_preset_pass_manager(1, backend, seed_transpiler=42)
pm.draw()

The flow consists of six stages:

print(pm.stages)

('init', 'layout', 'routing', 'translation', 'optimization', 'scheduling')

3.2 Draw an individual stage

First, let's draw all the tasks (transpiler passes) done in the init stage.

pm.init.draw()

We can run each individual stage. Let's run init stage for our circuit. By enabling logger, we can see the details of the run.

import logging

logger = logging.getLogger()
logger.setLevel("INFO")

init_out = pm.init.run(circuit)
init_out.draw("mpl", fold=-1)

INFO:qiskit.passmanager.base_tasks:Pass: UnitarySynthesis - 0.03576 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: HighLevelSynthesis - 0.16618 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: BasisTranslator - 0.07176 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: InverseCancellation - 0.27299 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: ContractIdleWiresInControlFlow - 0.00811 (ms)

3.3 Exercise

Draw layout stage passes and run the stage for the output circuit of the init stage (init_out), by modifying cells used above.

Solution:

display(pm.layout.draw())
layout_out = pm.layout.run(init_out)
layout_out.draw("mpl", idle_wires=False, fold=-1)

INFO:qiskit.passmanager.base_tasks:Pass: SetLayout - 0.01001 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: TrivialLayout - 0.07129 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: CheckMap - 0.08917 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: VF2Layout - 1.24431 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: BarrierBeforeFinalMeasurements - 0.02599 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: SabreLayout - 5.11169 (ms)

Do the same thing for translation stage.

Solution:

display(pm.translation.draw())
basis_out = pm.translation.run(layout_out)
basis_out.draw("mpl", idle_wires=False, fold=-1)

INFO:qiskit.passmanager.base_tasks:Pass: UnitarySynthesis - 0.03386 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: HighLevelSynthesis - 0.02718 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: BasisTranslator - 2.64192 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: CheckGateDirection - 0.02217 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: GateDirection - 0.36502 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: BasisTranslator - 0.64778 (ms)

Note: Any individual stage cannot always be run independently (as some of them need to carry over information from one previous stage).

3.4 Optimization Stage

The last default stage in the pipeline is optimization. After we've embedded the circuit for the target the circuit has expanded quite a bit. Most of this is due to inefficiencies in the equivalence relationships from basis translation and swap insertion. The optimization stage is used to try and minimize the size and depth of the circuit. It runs a series of passes in a do while loop until it reaches a steady output.

# pm.pre_optimization.draw()
pm.optimization.draw()

logger = logging.getLogger()
logger.setLevel("INFO")
opt_out = pm.optimization.run(basis_out)

INFO:qiskit.passmanager.base_tasks:Pass: Depth - 0.30112 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: FixedPoint - 0.03195 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: Size - 0.01216 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: FixedPoint - 0.01001 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: Optimize1qGatesDecomposition - 0.63729 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: InverseCancellation - 0.41723 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: ContractIdleWiresInControlFlow - 0.01192 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: GatesInBasis - 0.05484 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: Depth - 0.08583 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: FixedPoint - 0.20599 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: Size - 0.00787 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: FixedPoint - 0.00715 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: Optimize1qGatesDecomposition - 0.16809 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: InverseCancellation - 0.17190 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: ContractIdleWiresInControlFlow - 0.00691 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: GatesInBasis - 0.02408 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: Depth - 0.04935 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: FixedPoint - 0.00525 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: Size - 0.00620 (ms)
INFO:qiskit.passmanager.base_tasks:Pass: FixedPoint - 0.00286 (ms)

opt_out.draw("mpl", idle_wires=False, fold=-1)

4. In-depth examples

4.1 Two-qubit block optimization using two-qubit unitary synthesis

For level 2 and 3, we have more passes (Collect2qBlocks, ConsolidateBlocks, UnitarySynthesis) for more optimization, namely two-qubit block optimization. (Compare the optimization stage flow for level 2 with that above for level 1)

The two-qubit block optimization is composed of two steps: Collecting and consolidating 2-qubit blocks and synthesizing the 2-qubit unitary matrices.

pm2 = generate_preset_pass_manager(2, backend, seed_transpiler=42)
pm2.optimization.draw()

from qiskit.transpiler import PassManager
from qiskit.transpiler.passes import (
    Collect2qBlocks,
    ConsolidateBlocks,
    UnitarySynthesis,
)

# Collect 2q blocks and consolidate to unitary when we expect that we can reduce the 2q gate count for that unitary
consolidate_pm = PassManager(
    [
        Collect2qBlocks(),
        ConsolidateBlocks(target=backend.target),
    ]
)

display(basis_out.draw("mpl", idle_wires=False, fold=-1))

consolidated = consolidate_pm.run(basis_out)
consolidated.draw("mpl", idle_wires=False, fold=-1)

# Synthesize unitaries
UnitarySynthesis(target=backend.target)(consolidated).draw(
    "mpl", idle_wires=False, fold=-1
)

logger.setLevel("WARNING")

We saw in Part 2 that the real quantum compiler flow is not that simple and is composed of many passes (tasks). This is mainly due to the software engineering required to ensure performance for a wide range of application circuits and maintainability of the software. Qiskit transpiler would work well in most cases but if you happen to see your circuit is not well optimized by Qiskit transpiler, it would be a good opportunity to research your own application-specific circuit optimization as shown in Part 1. Transpiler technology is evolving, your R&D contribution is welcome.

from qiskit.circuit import QuantumCircuit
from qiskit_ibm_runtime import QiskitRuntimeService, Sampler

service = QiskitRuntimeService()
backend = service.backend("ibm_brisbane")
sampler = Sampler(backend)

circ = QuantumCircuit(3)
circ.ccx(0, 1, 2)
circ.measure_all()
circ.draw("mpl")

sampler.run([circ])  # IBMInputValueError will be raised

4.2 Circuit optimization matters

We first compare the results of running 5-qubit GHZ state ($\frac{1}{\sqrt{2}} \left( |00000\rangle + |11111\rangle \right)$) preparation circuits with and without optimization.

from qiskit.circuit import QuantumCircuit
from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager
from qiskit_ibm_runtime import QiskitRuntimeService, Sampler

service = QiskitRuntimeService()

# backend = service.backend('ibm_brisbane')
backend = service.least_busy(
    operational=True, simulator=False, min_num_qubits=127
)  # Eagle
backend

We first use a GHZ circuit naively synthesized as follows.

num_qubits = 5

ghz_circ = QuantumCircuit(num_qubits)
ghz_circ.h(0)
[ghz_circ.cx(0, i) for i in range(1, num_qubits)]
ghz_circ.measure_all()
ghz_circ.draw("mpl")

We transpile the circuit without optimization (optimization_level=0) and with optimization (optimization_level=2). As you can see, there is a big difference in the circuit length of transpiled circuits.

pm0 = generate_preset_pass_manager(
    optimization_level=0, backend=backend, seed_transpiler=777
)
pm2 = generate_preset_pass_manager(
    optimization_level=2, backend=backend, seed_transpiler=777
)
circ0 = pm0.run(ghz_circ)
circ2 = pm2.run(ghz_circ)
print("optimization_level=0:")
display(circ0.draw("mpl", idle_wires=False, fold=-1))
print("optimization_level=2:")
display(circ2.draw("mpl", idle_wires=False, fold=-1))

optimization_level=0:

optimization_level=2:

# run the circuits
sampler = Sampler(backend)
job = sampler.run([circ0, circ2], shots=10000)
job_id = job.job_id()
print(f"Job ID: {job_id}")

Job ID: d13rnnemya70008ek1zg

# REPLACE WITH YOUR OWN JOB IDS
job = service.job(job_id)

# get results
result = job.result()
unoptimized_result = result[0].data.meas.get_counts()
optimized_result = result[1].data.meas.get_counts()

from qiskit.visualization import plot_histogram

# plot
sim_result = {"0" * 5: 0.5, "1" * 5: 0.5}
plot_histogram(
    [result for result in [sim_result, unoptimized_result, optimized_result]],
    bar_labels=False,
    legend=[
        "ideal",
        "no optimization",
        "with optimization",
    ],
)

4.3 Circuit synthesis matters

We next compare the results of running two differently synthesized 5-qubit GHZ state ($\frac{1}{\sqrt{2}} \left( |00000\rangle + |11111\rangle \right)$) preparation circuits.

# Original GHZ circuit (naive synthesis)
ghz_circ.draw("mpl")

# A better GHZ circuit (smarter synthesis), you learned in a previous lecture
ghz_circ2 = QuantumCircuit(5)
ghz_circ2.h(2)
ghz_circ2.cx(2, 1)
ghz_circ2.cx(2, 3)
ghz_circ2.cx(1, 0)
ghz_circ2.cx(3, 4)
ghz_circ2.measure_all()
ghz_circ2.draw("mpl")

circ_org = pm2.run(ghz_circ)
circ_new = pm2.run(ghz_circ2)
print("original synthesis:")
display(circ_org.draw("mpl", idle_wires=False, fold=-1))
print("new synthesis:")
display(circ_new.draw("mpl", idle_wires=False, fold=-1))

original synthesis:

new synthesis:

# run the circuits
sampler = Sampler(backend)
job = sampler.run([circ_org, circ_new], shots=10000)
job_id = job.job_id()
print(f"Job ID: {job_id}")

Job ID: d13rp283grvg008j12fg

# REPLACE WITH YOUR OWN JOB IDS
job = service.job(job_id)

# get results
result = job.result()
synthesis_org_result = result[0].data.meas.get_counts()
synthesis_new_result = result[1].data.meas.get_counts()

# plot
sim_result = {"0" * 5: 0.5, "1" * 5: 0.5}
plot_histogram(
    [result for result in [sim_result, synthesis_org_result, synthesis_new_result]],
    bar_labels=False,
    legend=[
        "ideal",
        "synthesis_org",
        "synthesis_new",
    ],
)

4.4 General 1-qubit gate decomposition

from qiskit import QuantumCircuit, transpile
from qiskit.circuit import Parameter
from qiskit.circuit.library.standard_gates import UGate

phi, theta, lam = Parameter("φ"), Parameter("θ"), Parameter("λ")

qc = QuantumCircuit(1)
qc.append(UGate(theta, phi, lam), [0])
qc.draw(output="mpl")

transpile(qc, basis_gates=["rz", "sx"]).draw(output="mpl")

4.5 One-qubit block optimization

from qiskit import QuantumCircuit

qc = QuantumCircuit(1)
qc.x(0)
qc.y(0)
qc.z(0)
qc.rx(1.23, 0)
qc.ry(1.23, 0)
qc.rz(1.23, 0)
qc.h(0)
qc.s(0)
qc.t(0)
qc.sx(0)
qc.sdg(0)
qc.tdg(0)
qc.draw(output="mpl")

from qiskit.quantum_info import Operator

Operator(qc)

Operator([[ 0.45292511-0.57266982j, -0.66852684-0.14135058j],
          [ 0.14135058+0.66852684j, -0.57266982+0.45292511j]],
         input_dims=(2,), output_dims=(2,))

from qiskit import transpile

qc_opt = transpile(qc, basis_gates=["rz", "sx"])
qc_opt.draw(output="mpl")

Operator(qc_opt)

Operator([[ 0.45292511-0.57266982j, -0.66852684-0.14135058j],
          [ 0.14135058+0.66852684j, -0.57266982+0.45292511j]],
         input_dims=(2,), output_dims=(2,))

Operator(qc).equiv(Operator(qc_opt))

True

4.6 Toffoli decomposition

qc = QuantumCircuit(3)
qc.ccx(0, 1, 2)
qc.draw(output="mpl")

from qiskit import QuantumCircuit, transpile

qc = QuantumCircuit(3)
qc.ccx(0, 1, 2)
qc = transpile(qc, basis_gates=["rz", "sx", "cx"])
qc.draw(output="mpl")

4.7 CU gate decomposition

from qiskit.circuit.library.standard_gates import CUGate

phi, theta, lam, gamma = Parameter("φ"), Parameter("θ"), Parameter("λ"), Parameter("γ")
qc = QuantumCircuit(2)
# qc.cu(theta, phi, lam, gamma, 0, 1)
qc.append(CUGate(theta, phi, lam, gamma), [0, 1])
qc.draw(output="mpl")

from qiskit.circuit.library.standard_gates import CUGate

phi, theta, lam, gamma = Parameter("φ"), Parameter("θ"), Parameter("λ"), Parameter("γ")
qc = QuantumCircuit(2)
qc.append(CUGate(theta, phi, lam, gamma), [0, 1])
qc = transpile(qc, basis_gates=["rz", "sx", "cx"])
qc.draw(output="mpl")

4.8 CX, ECR, CZ equal up to local Cliffords

Note that $H$(Hadamard), $S$($\pi/2$ Z-rotation), $S^\dagger$($-\pi/2$ Z-rotation), $X$(Pauli X) are all Clifford gates.

qc = QuantumCircuit(2)
qc.cx(0, 1)
qc.draw(output="mpl", style="bw")

qc = QuantumCircuit(2)
qc.cx(0, 1)
transpile(qc, basis_gates=["x", "s", "h", "sdg", "ecr"]).draw(output="mpl", style="bw")

qc = QuantumCircuit(2)
qc.cx(0, 1)
transpile(qc, basis_gates=["h", "cz"]).draw(output="mpl", style="bw")

Using IBM backend 1q basis gates "rz", "sx" and "x".

qc = QuantumCircuit(2)
qc.cx(0, 1)
transpile(qc, basis_gates=["rz", "sx", "x", "ecr"]).draw(output="mpl", style="bw")

qc = QuantumCircuit(2)
qc.cx(0, 1)
transpile(qc, basis_gates=["rz", "sx", "x", "cz"]).draw(output="mpl", style="bw")

# Check Qiskit version
import qiskit

qiskit.__version__

'2.0.2'