Implementing runtime parameter binding

[doichanj]

Summary

This is new feature to optimize GPU simulation for single circuit with multiple parameters.
Before this PR, multiple circuits were given by binding parameters from a single circuit. This PR bind parameters to each gates at runtime on a single circuit with multiple shots of simulations, so we can speed up by combining batched execution on GPU.

Details and comments

Runtime parameter binding puts list of parameters to each operation and parameter is selected for each shot at runtime.
This option changes parallelization from multi-experiments to multi-shots.

[Guogggg]

I downloaded your code, and compiled from qiskit-aer source, and set up

noiseless_estimator = AerEstimator(
    backend_options={"method": "statevector","device":"GPU","runtime_parameter_bind_enable":True,"batched_shots_gpu":True},
    run_options={"shots": None},
    approximation=True,
)

However, there is no significant acceleration in GPU operation time, so here is my code

import numpy as np
import sys

from qiskit_nature.units import DistanceUnit
from qiskit_nature.second_q.drivers import PySCFDriver
from qiskit_nature.second_q.mappers import JordanWignerMapper, ParityMapper
from qiskit.algorithms.minimum_eigensolvers import VQE
from qiskit.algorithms.minimum_eigensolvers import MinimumEigensolver
from qiskit.algorithms.optimizers import SLSQP
from qiskit.algorithms.optimizers import SPSA
from qiskit.algorithms.optimizers import L_BFGS_B

from qiskit.primitives import BackendEstimator
from qiskit_aer import AerSimulator
from qiskit_nature.second_q.circuit.library import HartreeFock, UCCSD
from qiskit_nature.second_q.algorithms import GroundStateEigensolver
import time
from qiskit_aer.primitives import Estimator as AerEstimator
from qiskit_nature.settings import  QiskitNatureSettings
QiskitNatureSettings.use_pauli_sum_op = False
print(sys.path)
dis = 1.5
driver = PySCFDriver(
        atom = "Li 0 0 0; H 0 0 " + str(dis),
        basis = "sto3g",
        charge = 0,
        spin = 0,
        unit = DistanceUnit.ANGSTROM,
    )
     
es_problem = driver.run()

mapper_jw = ParityMapper()

ansatz = UCCSD(es_problem.num_spatial_orbitals,
		es_problem.num_particles,
                mapper_jw,
                initial_state=HartreeFock(es_problem.num_spatial_orbitals,
                                           es_problem.num_particles,
                                           mapper_jw),
                )

noiseless_estimator = AerEstimator(
    backend_options={"method": "statevector","device":"GPU","runtime_parameter_bind_enable":True,"batched_shots_gpu":True},
    run_options={"shots": None},
    approximation=True,
)
vqe_solver = VQE(
    noiseless_estimator, ansatz, optimizer=SLSQP()
)

vqe_solver.initial_point = [0] * ansatz.num_parameters
st = time.time()
fermionic_op = es_problem.hamiltonian.second_q_op()
qubit_op = mapper_jw.map(fermionic_op)    # 12-qubit
result = vqe_solver.compute_minimum_eigenvalue(qubit_op)

calc = GroundStateEigensolver(mapper_jw, vqe_solver)
res = calc.solve(es_problem)
energy = result.eigenvalue + res.nuclear_repulsion_energy
print(energy)
print("time used: ",time.time()-st)

@doichanj

[doichanj]

I tested script above and time comparison is as followings;

• CPU w/o runtime parameter binding : 454.241402387619
• CPU w/runtime parameter binding : 476.4489688873291
• GPU w/o runtime parameter binding : 380.32021927833557
• GPU w/runtime parameter binding : 359.3335130214691

I observed small speedup by using runtime parameter binding on GPU, but I think most of the time is spent outside of simulator in this case.

[Guogggg]

I tested script above and time comparison is as followings;

• CPU w/o runtime parameter binding : 454.241402387619
• CPU w/runtime parameter binding : 476.4489688873291
• GPU w/o runtime parameter binding : 380.32021927833557
• GPU w/runtime parameter binding : 359.3335130214691

I observed small speedup by using runtime parameter binding on GPU, but I think most of the time is spent outside of simulator in this case.
Is there some other way we can make compute_minimum_eigenvalue faster, and I think it's not parallel, I set "max_parallel_threads":14,"statevector_parallel_threshold":6 in the source code, so that it takes more core, but the computation time is longer

[hhorii]

I think a major change to the existing code is introducing ResultItr and it looks reasonable. Though I was not able to check the codes completely, added/modified codes look good. If they passed all the existing codes, merging this PR has no problem (assuming fixing bugs when they are found).

I left some very minor comments. If they are fixed, this PR looks good to be merged.

[hhorii]

Is this necessary?

[doichanj]

removed

[hhorii]

binded -> bound because other descriptions use the latter in Aer.

[doichanj]

fixed

[hhorii]

I think this method name is bind_parameters().

[doichanj]

changed name to bind_parameter because this function bind single parameter