Improve performance of RB circuit generation

[eendebakpt]

Summary

The circuit generation for RB experiments is slow due to the use of indexing on sparse matrices (_CLIFFORD_COMPOSE_2Q from qiskit_experiments.library.randomized_benchmarking.clifford_utils). Since indexing is the only operation used on _CLIFFORD_COMPOSE_2Q we can improve performance by using the lil_matrix format.

Details and comments

Improve performance of RB circuit generation by changing the format of _CLIFFORD_COMPOSE_2Q to lil_matrix.

Micro benchmark:

import scipy.sparse
from qiskit_experiments.library.randomized_benchmarking.clifford_utils import _CLIFFORD_COMPOSE_2Q

_CLIFFORD_COMPOSE_2Q_lil =  scipy.sparse.lil_matrix(_CLIFFORD_COMPOSE_2Q)

%timeit _CLIFFORD_COMPOSE_2Q[2,3]
%timeit _CLIFFORD_COMPOSE_2Q_lil[2,3]

Result:

29.4 µs ± 4.01 µs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)
2.64 µs ± 777 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

Application benchmark:

import time
import qiskit
from qiskit_experiments.library import InterleavedRB

gate=qiskit.circuit.library.CXGate()
rb=InterleavedRB(gate, physical_qubits= (0,1),lengths= [1,10,50,100,200], num_samples= 10)

t0=time.time()
c=rb.circuits()
dt=time.time()-t0

results in a speedup from 2.1419544219970703 to 1.154689073562622 seconds.

PR checklist (delete when all criteria are met)

• I have read the contributing guide CONTRIBUTING.md.
• I have added the tests to cover my changes.
• I have updated the documentation accordingly.
• I have added a release note file using reno if this change needs to be documented in the release notes. (no release note required)

[itoko]

Thanks, @eendebakpt I think this is really a nice idea. Although it requires a matrix conversion once when loading the data, I've confirmed the conversion overhead is negligible (less than 0.1 seconds in my env).
I approve as the test failure seems irrelevant to the change in this PR.