cirq.resolve_parameters is slow

[rht]

Description of the issue

The usage context is to run hybrid variational circuit fast. For a small circuit of 2 qubits, consisting of ZZFeatureMap and RealAmplitude, I find cirq.resolve_parameters to be ~15x slower than if I were to reconstruct the circuit from scratch with concrete numbers replacing the parameters instead.

How to reproduce the issue

import time

import numpy as np
import cirq
import sympy


def make_cirq_circuit(
    num_qubits,
    circuit_params,
    qubits,
    cirq_input_params,
    cirq_weight_params,
):
    circuit = cirq.Circuit()
    # Qiskit's ZZFeatureMap
    zz_repetitions = 2
    for i in range(zz_repetitions):
        circuit.append([cirq.H(q) for q in qubits])
        circuit.append(
            [cirq.Z(qubits[i]) ** (2.0 * cirq_input_params[i] / np.pi)]
            for i in range(num_qubits)
        )
        circuit.append(cirq.CX(qubits[0], qubits[1]))
        circuit.append(
            cirq.Z(qubits[1])
            ** (
                (2.0 * (np.pi - cirq_input_params[0]) * (np.pi - cirq_input_params[1]))
                / np.pi
            )
        )
    # Qiskit's RealAmplitude
    count = 0
    for i in range(4):
        circuit.append(cirq.CX(qubits[0], qubits[1]))
        circuit.append(cirq.ry(cirq_weight_params[count * 2]).on(qubits[0]))
        circuit.append(cirq.ry(cirq_weight_params[count * 2 + 1]).on(qubits[1]))
        count += 1
    return circuit


num_qubits = 2
qubits = cirq.LineQubit.range(num_qubits)

# Prepare the parameterized circuit
cirq_input_params = [sympy.Symbol("x0"), sympy.Symbol("x1")]
cirq_weight_params = [sympy.Symbol(f"theta{i}") for i in range(8)]
circuit_params = cirq_input_params + cirq_weight_params
circuit = make_cirq_circuit(
    num_qubits,
    circuit_params,
    qubits,
    cirq_input_params,
    cirq_weight_params,
)

# Resolve params
np.random.seed(42)
p = {p: np.random.random() for p in circuit_params}
tic = time.time()
resolved_circuit = cirq.resolve_parameters(circuit, p)
print("Elapsed cirq.resolve_parameters:", round(time.time() - tic, 3))

# What if we instead construct the circuit from scratch.
cirq_input_params = [np.random.random() for i in range(2)]
cirq_weight_params = [np.random.random() for i in range(8)]
circuit_params = cirq_input_params + cirq_weight_params
tic = time.time()
make_cirq_circuit(
    num_qubits,
    circuit_params,
    qubits,
    cirq_input_params,
    cirq_weight_params,
)
print("Elapsed construct from scratch:", round(time.time() - tic, 3))

Result:

Elapsed cirq.resolve_parameters: 0.045
Elapsed construct from scratch: 0.003

Cirq version
You can get the cirq version by printing cirq.__version__. From the command line:

1.1.0

[rht]

Past issue: #1899 and its resolution #3205.

With Qiskit:

from qiskit import QuantumCircuit
from qiskit.circuit.library import RealAmplitudes, ZZFeatureMap
feature_map = ZZFeatureMap(2, reps=2)
ansatz = RealAmplitudes(2)
qc = QuantumCircuit(2)
qc.compose(feature_map, inplace=True)
qc.compose(ansatz, inplace=True)
tic = time.time()
qc.bind_parameters(p)
print("Elapsed Qiskit:", round(time.time() - tic, 3))
# Elapsed Qiskit: 0.007

[rht]

Tracking down the slowest part of cirq.resolve_parameters for this example, >90% of the time is spent on

Cirq/cirq-core/cirq/protocols/resolve_parameters.py

Line 189 in a95f009

result = getter(param_resolver, recursive)

, and in the ParamResolver class, >90% of the time is spent in this line:

Cirq/cirq-core/cirq/study/resolver.py

Line 187 in a95f009

v = value.subs(self.param_dict, simultaneous=True)

.

[rht]

Replacing import sympy with import symengine as sympy throughout the library, I got

Elapsed cirq.resolve_parameters: 0.011

[tanujkhattar]

@NoureldinYosri to test if its potentially a library loading issue by rerunning the test multiple times.

We should also look at symengine and see if it breaks any of our tests.

[NoureldinYosri]

well it's not a library loading issue.

[skushnir123]

I can take up this issue.

[skushnir123]

@rht I merged this PR #6128 a couple weeks back to improve performance quite a bit. Let me know if things look good on your end!