Convert ASV benchmarks to pytest-benchmark (#7799)

This duplicates ASV benchmarks as follows:

- `circuit_construction.py` to `circuit_construction_perf.py`
- `bench_linalg_decompositions.py` to `linalg_decompositions_perf.py`
- `parameter_resolution.py` to `parameter_resolution_perf.py`
- `randomized_benchmarking.py` to `randomized_benchmarking_perf.py`
- `serialization.py` to `serialization_perf.py`
- `transformers/routing.py` to `transformers/routing_perf.py`
- `transformers/transformer_primitives.py` to
`transformers/transformer_primitives_perf.py`

Partially implements #7796

Author: pavoljuhas

asv.conf.json

@@ -7,6 +7,7 @@
     "build_command": [
         "PIP_NO_BUILD_ISOLATION=false python -m pip wheel --no-deps --no-index -w {build_cache_dir} {build_dir}"
     ],
+    "install_command": ["in-dir={env_dir} python -m pip install {wheel_file} pytest"],
     "branches": ["main"],
     "dvcs": "git",
     "environment_type": "virtualenv",

benchmarks/circuit_construction_perf.py

@@ -0,0 +1,166 @@
+# Copyright 2026 The Cirq Developers
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Performance tests for circuit construction."""
+
+import itertools
+from collections.abc import Sequence
+
+import pandas
+import pytest
+
+import cirq
+
+
+def rotated_surface_code_memory_z_cycle(
+    data_qubits: set[cirq.GridQubit],
+    z_measure_qubits: set[cirq.GridQubit],
+    x_measure_qubits: set[cirq.GridQubit],
+    z_order: Sequence[tuple[int, int]],
+    x_order: Sequence[tuple[int, int]],
+) -> cirq.Circuit:
+    """Constructs a circuit for a single round of rotated memory Z surface code.
+
+    Args:
+        data_qubits: data qubits for the surface code patch.
+        z_measure_qubits: measure qubits to measure Z stabilizers for surface code patch.
+        x_measure_qubits: measure qubits to measure X stabilizers for surface code patch.
+        z_order: Specifies the order in which the 2/4 data qubit neighbours of a Z measure qubit
+            should be processed.
+        x_order: Specifies the order in which the 2/4 data qubit neighbours of a X measure qubit
+            should be processed.
+
+    Returns:
+        A `cirq.Circuit` for a single round of rotated memory Z surface code cycle.
+    """
+
+    circuit = cirq.Circuit()
+    circuit += cirq.Moment([cirq.H(q) for q in x_measure_qubits])
+    for k in range(4):
+        op_list = []
+        for measure_qubits, add, is_x in [
+            (x_measure_qubits, x_order[k], True),
+            (z_measure_qubits, z_order[k], False),
+        ]:
+            for q_meas in measure_qubits:
+                q_data = q_meas + add
+                if q_data in data_qubits:
+                    op_list.append(cirq.CNOT(q_meas, q_data) if is_x else cirq.CNOT(q_data, q_meas))
+        circuit += cirq.Moment(op_list)
+    circuit += cirq.Moment([cirq.H(q) for q in x_measure_qubits])
+    circuit += cirq.Moment(cirq.measure_each(*x_measure_qubits, *z_measure_qubits))
+    return circuit
+
+
+def surface_code_circuit(
+    distance: int, num_rounds: int, moment_by_moment: bool = True
+) -> cirq.Circuit:
+    """Constructs a rotated memory Z surface code circuit with `distance` and `num_rounds`.
+
+    The circuit has `dxd` data qubits and `d ** 2 - 1` measure qubits, where `d` is the distance
+    of surface code. For more details on rotated surface codes and qubit indexing, see figure 13
+    https://arxiv.org/abs/1111.4022.
+
+    Args:
+        distance: Distance of the surface code.
+        num_rounds: Number of error correction rounds for memory Z experiment.
+        moment_by_moment: If True, the circuit is constructed moment-by-moment instead of
+            operation-by-operation. This is useful to benchmark different circuit construction
+            patterns for the same circuit.
+
+    Returns:
+        A `cirq.Circuit` for surface code memory Z experiment for `distance` and `num_rounds`.
+    """
+
+    def ndrange(*ranges: tuple[int, ...]):
+        return itertools.product(*[range(*r) for r in ranges])
+
+    data_qubits = {cirq.q(2 * x + 1, 2 * y + 1) for x, y in ndrange((distance,), (distance,))}
+    z_measure_qubits = {
+        cirq.q(2 * x, 2 * y) for x, y in ndrange((1, distance), (distance + 1,)) if x % 2 != y % 2
+    }
+    x_measure_qubits = {
+        cirq.q(2 * x, 2 * y) for x, y in ndrange((distance + 1,), (1, distance)) if x % 2 == y % 2
+    }
+    x_order = [(1, 1), (1, -1), (-1, 1), (-1, -1)]
+    z_order = [(1, 1), (-1, 1), (1, -1), (-1, -1)]
+    surface_code_cycle = rotated_surface_code_memory_z_cycle(
+        data_qubits, x_measure_qubits, z_measure_qubits, x_order, z_order
+    )
+    if moment_by_moment:
+        return cirq.Circuit(
+            surface_code_cycle * num_rounds, cirq.Moment(cirq.measure_each(*data_qubits))
+        )
+    else:
+        return cirq.Circuit(
+            [*surface_code_cycle.all_operations()] * num_rounds, cirq.measure_each(*data_qubits)
+        )
+
+
+class TestSurfaceCodeRotatedMemoryZ:
+    """Surface Code Rotated Memory-Z Benchmarks."""
+
+    group = "circuit_construction"
+    expected = pandas.DataFrame.from_dict(
+        {
+            3: [64, 369],
+            5: [176, 3225],
+            7: [344, 12985],
+            9: [568, 36369],
+            11: [848, 82401],
+            13: [1184, 162409],
+            15: [1576, 290025],
+            17: [2024, 481185],
+            19: [2528, 754129],
+            21: [3088, 1129401],
+            23: [3704, 1629849],
+            25: [4376, 2280625],
+        },
+        columns=["depth", "operation_count"],
+        orient="index",
+    )
+
+    @pytest.mark.parametrize("distance", expected.index)
+    @pytest.mark.benchmark(group=group)
+    def test_circuit_construction_moment_by_moment(self, benchmark, distance: int) -> None:
+        """Benchmark circuit construction for Rotated Bottom-Z Surface code."""
+        circuit = benchmark(
+            surface_code_circuit, distance, num_rounds=distance * distance, moment_by_moment=True
+        )
+        assert len(circuit) == self.expected.depth[distance]
+
+    @pytest.mark.parametrize("distance", expected.index)
+    @pytest.mark.benchmark(group=group)
+    def test_circuit_construction_operations_by_operation(self, benchmark, distance: int) -> None:
+        """Benchmark circuit construction for Rotated Bottom-Z Surface code."""
+        circuit = benchmark(
+            surface_code_circuit, distance, num_rounds=distance * distance, moment_by_moment=False
+        )
+        assert sum(1 for _ in circuit.all_operations()) == self.expected.operation_count[distance]
+
+
+class TestXOnAllQubitsCircuit:
+    """N * D times X gate on all qubits."""
+
+    group = "circuit_operations"
+
+    @pytest.mark.parametrize(
+        ["qubit_count", "depth"], itertools.product([1, 10, 100, 1000], [1, 10, 100, 1000])
+    )
+    @pytest.mark.benchmark(group=group)
+    def test_circuit_construction(self, benchmark, qubit_count: int, depth: int) -> None:
+        q = cirq.LineQubit.range(qubit_count)
+        f = lambda: cirq.Circuit(cirq.Moment(cirq.X.on_each(*q)) for _ in range(depth))
+        circuit = benchmark(f)
+        assert len(circuit) == depth

benchmarks/linalg_decompositions_perf.py

@@ -0,0 +1,26 @@
+# Copyright 2026 The Cirq Developers
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import pytest
+
+import cirq
+
+
+@pytest.mark.parametrize(
+    "gate", [cirq.IdentityGate(2), cirq.SWAP, cirq.ISWAP, cirq.CZ, cirq.CNOT], ids=str
+)
+@pytest.mark.benchmark(group="linalg_decompositions")
+def test_kak_decomposition(benchmark, gate: cirq.Gate) -> None:
+    """Benchmark kak_decomposition."""
+    benchmark(cirq.kak_decomposition, gate)

benchmarks/parameter_resolution_perf.py

@@ -0,0 +1,41 @@
+# Copyright 2026 The Cirq Developers
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import itertools
+import random
+
+import numpy as np
+import pytest
+import sympy
+
+import cirq
+
+
+@pytest.mark.parametrize(
+    ["num_qubits", "num_scan_points"], itertools.product([50, 100, 150, 200], [20, 40, 60, 80, 100])
+)
+@pytest.mark.benchmark(group="parameter_resolution")
+def test_parameter_resolution(benchmark, num_qubits: int, num_scan_points: int) -> None:
+    qubits = cirq.GridQubit.rect(1, num_qubits)
+    symbols = {q: sympy.Symbol(f'a_{q}') for q in qubits}
+    circuit = cirq.Circuit([cirq.X(q) ** symbols[q] for q in qubits], cirq.measure_each(*qubits))
+    qubit_amps = {q: random.uniform(0.48, 0.52) for q in qubits}
+    diff_amps = np.linspace(-0.3, 0.3, num=num_scan_points)
+
+    def _f():
+        for diff in diff_amps:
+            resolver = {symbols[q]: amp + diff for q, amp in qubit_amps.items()}
+            _ = cirq.resolve_parameters(circuit, resolver)
+
+    benchmark(_f)

benchmarks/randomized_benchmarking_perf.py

@@ -0,0 +1,101 @@
+# Copyright 2026 The Cirq Developers
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import functools
+import itertools
+from collections.abc import Iterable, Sequence
+
+import numpy as np
+import pytest
+
+import cirq
+from cirq.experiments.qubit_characterizations import _find_inv_matrix, _single_qubit_cliffords
+
+
+def dot(args: Iterable[np.ndarray]) -> np.ndarray:
+    return functools.reduce(np.dot, args)
+
+
+class TestSingleQubitRandomizedBenchmarking:
+    """Benchmarks circuit construction time for single qubit randomized benchmarking circuits.
+
+    Given a combination of `depth`, `num_qubits` and `num_circuits`, the benchmark constructs
+    `num_circuits` different circuits, each spanning `num_qubits` and containing `depth` moments.
+    Each moment of the circuit contains a single qubit clifford operation for each qubit.
+
+    Thus, the generated circuits have `depth * num_qubits` single qubit clifford operations.
+    """
+
+    group = "randomized_benchmarking"
+    depth = [1, 10, 50, 100, 250, 500, 1000]
+    num_qubits = [100]
+    num_circuits = [20]
+
+    # assigned in setup_class
+    sq_xz_matrices: np.ndarray
+    sq_xz_cliffords: Sequence[cirq.Gate]
+
+    @classmethod
+    def setup_class(cls):
+        cls.sq_xz_matrices = np.array(
+            [
+                dot([cirq.unitary(c) for c in reversed(group)])
+                for group in _single_qubit_cliffords().c1_in_xz
+            ]
+        )
+        cls.sq_xz_cliffords = [cirq.PhasedXZGate.from_matrix(mat) for mat in cls.sq_xz_matrices]
+
+    def _get_op_grid(self, qubits: Sequence[cirq.Qid], depth: int) -> list[list[cirq.Operation]]:
+        op_grid: list[list[cirq.Operation]] = []
+        for q in qubits:
+            gate_ids = np.random.choice(len(self.sq_xz_cliffords), depth)
+            idx = _find_inv_matrix(dot(self.sq_xz_matrices[gate_ids][::-1]), self.sq_xz_matrices)
+            op_sequence = [self.sq_xz_cliffords[gate_id].on(q) for gate_id in gate_ids]
+            op_sequence.append(self.sq_xz_cliffords[idx].on(q))
+            op_grid.append(op_sequence)
+        return op_grid
+
+    @pytest.mark.parametrize(
+        ["depth", "num_qubits", "num_circuits"], itertools.product(depth, num_qubits, num_circuits)
+    )
+    @pytest.mark.benchmark(group=group)
+    def test_rb_op_grid_generation(
+        self, benchmark, depth: int, num_qubits: int, num_circuits: int
+    ) -> None:
+        qubits = cirq.GridQubit.rect(1, num_qubits)
+
+        def _f() -> None:
+            for _ in range(num_circuits):
+                self._get_op_grid(qubits, depth)
+
+        benchmark(_f)
+
+    @pytest.mark.parametrize(
+        ["depth", "num_qubits", "num_circuits"], itertools.product(depth, num_qubits, num_circuits)
+    )
+    @pytest.mark.benchmark(group=group)
+    def test_rb_circuit_construction(
+        self, benchmark, depth: int, num_qubits: int, num_circuits: int
+    ) -> None:
+        qubits = cirq.GridQubit.rect(1, num_qubits)
+
+        def _f() -> None:
+            for _ in range(num_circuits):
+                op_grid = self._get_op_grid(qubits, depth)
+                cirq.Circuit(
+                    [cirq.Moment(ops[d] for ops in op_grid) for d in range(depth + 1)],
+                    cirq.Moment(cirq.measure(*qubits)),
+                )
+
+        benchmark(_f)