Implements optimizations A.1 and A.2 for Quantum Shannon Decomposition

[codrut3]

Fixes #6777

I checked that with the current PR, the circuit depth after transpiler matches that obtained by Qiskit.

The runtime is slow due merge_single_qubit_gates_to_phased_x_and_z. I wrote a different implementation for the special case used by the Shannon decomposition. With this, the runtime is halved. I believe merge_single_qubit_gates_to_phased_x_and_z can be optimized too.

[codrut3]

@NoureldinYosri Please have a look!

[codrut3]

For future reference, this is the script I used to benchmark the decomposition:

# pylint: disable=wrong-or-nonexistent-copyright-notice
"""Tests performance of quantum Shannon decomposition.
"""

import sys
import time

from scipy.stats import unitary_group

import cirq
import qiskit.qasm2


if not sys.warnoptions:
    import warnings
    warnings.simplefilter("ignore")


def main():
    for n_qubits in range(3, 9):
        U = unitary_group.rvs(2**n_qubits)
        qubits = [cirq.NamedQubit(f'q{i}') for i in range(n_qubits)]
        start_time = time.time()
        circuit = cirq.Circuit(cirq.quantum_shannon_decomposition(qubits, U))
        elapsed_time = time.time() - start_time
        print(f'Depth for {n_qubits} qubits (no transpiler): {len(circuit)} ({elapsed_time:.2f} sec)')
        print(f'Number of CZ gates (no transpiler): {len(list(circuit.findall_operations_with_gate_type(cirq.CZPowGate)))}')
        print(f'Number of CNOT gates (no transpiler): {len(list(circuit.findall_operations_with_gate_type(cirq.CXPowGate)))}')


        # Export to qasm
        qasm_str = cirq.qasm(circuit, args=cirq.QasmArgs(version="2.0"))
        qiskit_circuit = qiskit.qasm2.loads(qasm_str, include_path=('.',), custom_instructions=qiskit.qasm2.LEGACY_CUSTOM_INSTRUCTIONS, custom_classical=(), strict=False)

        # Run qiskit transpiler
        start_time = time.time()
        opt_circuit = qiskit.compiler.transpile(qiskit_circuit, basis_gates=["cz", "u3"], optimization_level=3)
        elapsed_time = time.time() - start_time
        print(f'Qiskit depth for {n_qubits} qubits (with transpiler): {opt_circuit.depth()} ({elapsed_time:.2f} sec)')
        print(f'Qiskit gate counts for {n_qubits} qubits (with transpiler): {dict(opt_circuit.count_ops())}')

        # Run cirq transpiler
        start_time = time.time()
        circuit = cirq.optimize_for_target_gateset(circuit, gateset=cirq.CZTargetGateset(preserve_moment_structure=False), max_num_passes=None)
        elapsed_time = time.time() - start_time
        print(f'Cirq depth for {n_qubits} qubits (with transpiler): {len(circuit)} ({elapsed_time:.2f} sec)')
        # Count the operations
        operation_counts = {}
        for op in circuit.all_operations():
            gate_type = op.gate.__class__
            if isinstance(op.gate, cirq.CZPowGate):
                gate_type = 'cz'
            elif isinstance(op.gate, cirq.PhasedXZGate):
                gate_type = 'u3'
            if gate_type not in operation_counts:
                operation_counts[gate_type] = 0
            operation_counts[gate_type] += 1
        print(f'Cirq gate counts for {n_qubits} qubits (with transpiler): {operation_counts}')


if __name__ == '__main__':
    main()

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 98.69%. Comparing base (12eab31) to head (f44f328).
Report is 26 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #7390      +/-   ##
==========================================
+ Coverage   98.68%   98.69%   +0.01%     
==========================================
  Files        1112     1112              
  Lines       97642    98027     +385     
==========================================
+ Hits        96359    96749     +390     
+ Misses       1283     1278       -5     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[NoureldinYosri]

thanks @codrut3 and sorry for the late review ... I started reviewing and forgot to send

[NoureldinYosri]

why MatrixGate in particular? shouldn't is op.gate is not None be sufficient

[codrut3]

I need a way to differentiate between end gates produced by the Shannon decomposition (such as CZ, CNOT), and 2-qubit gates that need to be further decomposed in the A.2 optimization step. A.2 requires splicing a diagonal matrix from the 2-gate decomposition and multiplying it with the previous 2-qubit gate. So I extract these gates in a separate list, perform A.2, and then replace the gates in the original list with the decomposition result.

[NoureldinYosri]

we don't use SimpleNamespace anywhere else in Cirq. Also SimpleNamespace doesn't support type checking. can you replace it with an attrs dataclass, maybe even attrs.frozen?

[codrut3]

Thank you, I replaced with attrs.define. I can't use frozen because the matrix gets updated during the A.2 step: it gets multiplied with the diagonal gate extracted from the next gate decomposition.

If the Shannon decomposition is something like:
2-gate A, 1-qubit gate, 1-qubit gate, 2-gate B, ...
then for A.2 first I decompose B into a diagonal gate and a sequence of operations with at most 2 CNOTs:
2-gate A, 1-qubit gate, 1-qubit gate, (D, [sequence of operations]), ...
then multiply D (a diagonal matrix) with A:
D x 2-gate A, 1-qubit gate, 1-qubit gate, [sequence of operations]), ...
then decompose D x A, and so on.

[NoureldinYosri]

I know you didn't write this but can you modify this line to be

Suggested change

	if np.abs(global_phase) > 1e-9:
	if not np.isclose(global_phase, 0, atol=atol):

[codrut3]

Done.

[NoureldinYosri]

Suggested change

	for op in shannon_decomp:
	yield from ops.flatten_op_tree(shannon_decomp)

[codrut3]

Done. Unfortunately mypy required a cast.

[NoureldinYosri]

do we still need this? I thought that the analytical decompositions now preserve global phase #6523, is this not the case?

[codrut3]

I tried to remove it, but unit tests are failing. Looking at #6523 , it seems to me it was mistakenly closed by #7118 . #7118 changes the MatrixGate decomposition, but #6523 requires updating two_qubit_to_cz, a different thing.

[NoureldinYosri]

instead of passing an arbitary atol below, take it as input

Suggested change

	) -> Sequence[ops.Operation]:
	atol: float,
	) -> Sequence[ops.Operation]:

[codrut3]

Done.

[NoureldinYosri]

Suggested change

	for g in single_qubit_decompositions.single_qubit_matrix_to_phased_x_z(u, atol=1e-8)
	for g in single_qubit_decompositions.single_qubit_matrix_to_phased_x_z(u, atol=atol)

[codrut3]

Done.

[NoureldinYosri]

can you make this a public method in a seprate PR, seems like a useful transformation?

[codrut3]

I would like to optimize merge_single_qubit_gates_to_phased_x_and_z in another PR. Actually I wrote this implementation after reading the code for merge_single_qubit_gates_to_phased_x_and_z, just without the many conversions to frozen circuit :)

[codrut3]

Thank you a lot for the comments @NoureldinYosri ! PTAL