MattePalte
commented
2 years ago The problem seems related to the transformation to qasm format (qc.qasm()), because the two machine use different decomposers TwoQubitBasisDecomposer, which in turn come from two different TwoQubitWeylDecomposition.
The TwoQubitBasisDecomposer is initialized here in my qiskit version:
https://github.com/Qiskit/qiskit-terra/blob/a8a0d1be26b5df097e7a21016fa21550a3c4efce/qiskit/quantum_info/synthesis/two_qubit_decompose.py#L1405
But it is similar in the latest version: https://github.com/Qiskit/qiskit-terra/blob/fe8ab43f1218d6997458fb3bf4011045f40932b0/qiskit/quantum_info/synthesis/two_qubit_decompose.py#L1449 https://github.com/Qiskit/qiskit-terra/blob/fe8ab43f1218d6997458fb3bf4011045f40932b0/qiskit/quantum_info/synthesis/two_qubit_decompose.py#L1422
The difference can be exposed with this even minimal example.
from qiskit.quantum_info.synthesis.two_qubit_decompose import TwoQubitWeylDecomposition
from qiskit.quantum_info.operators import Operator
from qiskit.circuit.library.standard_gates import CXGate
gate = CXGate()
weyl = TwoQubitWeylDecomposition(Operator(gate).data)
weyl.__dict__Output SETUP A:
{'_original_decomposition': TwoQubitWeylDecomposition.from_bytes(
# TwoQubitWeylDecomposition(
# global phase: π/4
# ┌──────────┐┌──────────┐ ┌────────────┐┌────────┐┌──────────┐
# q_0: ┤ Rz(-π/2) ├┤ Ry(-π/2) ├────────────┤0 ├┤ Rz(-π) ├┤ Ry(-π/2) ├
# ├──────────┤├─────────┬┘┌──────────┐│ Rxx(-π/2) │├───────┬┘└──────────┘
# q_1: ┤ Rz(-π/2) ├┤ Ry(π/2) ├─┤ Rz(-π/2) ├┤1 ├┤ Ry(π) ├─────────────
# └──────────┘└─────────┘ └──────────┘└────────────┘└───────┘
# )
b'k05VTVBZAQB2AHsnZGVzY3InOiAnPGMxNicsICdmb3J0cmFuX29yZGVyJzogRmFsc2UsICdzaGFw'
b'ZSc6ICg0LCA0KSwgfSAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAg'
b'ICAgICAgICAgICAgIAoAAAAAAADwPwAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA'
b'AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA'
b'AAAAAAAAAAAAAAAAAAAAAAAA8D8AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA'
b'AAAAAAAAAAAA8D8AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA'
b'8D8AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA',
requested_fidelity=0.999999999,
calculated_fidelity=1.0,
actual_fidelity=0.9999999999999997,
abc=(0.7853981633974483, 0.0, 0.0)),
'a': 0.7853981633974483,
'b': 0,
'c': 0,
'K1l': array([[-1.05226633e-16+3.82683432e-01j, 9.23879533e-01-5.36741351e-17j],
[-9.23879533e-01+6.41070206e-18j, 3.98309960e-17-3.82683432e-01j]]),
'K1r': array([[-0.5-0.5j, 0.5+0.5j],
[-0.5+0.5j, -0.5+0.5j]]),
'K2l': array([[ 5.65713056e-17+3.82683432e-01j, -9.23879533e-01-2.34326020e-17j],
[ 9.23879533e-01-2.34326020e-17j, 5.65713056e-17-3.82683432e-01j]]),
'K2r': array([[ 4.32978028e-17-0.70710678j, -4.32978028e-17+0.70710678j],
[ 4.32978028e-17+0.70710678j, 4.32978028e-17+0.70710678j]]),
'global_phase': -2.356194490192345,
'unitary_matrix': array([[1.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[0.+0.j, 0.+0.j, 0.+0.j, 1.+0.j],
[0.+0.j, 0.+0.j, 1.+0.j, 0.+0.j],
[0.+0.j, 1.+0.j, 0.+0.j, 0.+0.j]]),
'requested_fidelity': None,
'_is_flipped_from_original': False,
'calculated_fidelity': 1.0}Output SETUP B:
{'_original_decomposition': TwoQubitWeylDecomposition.from_bytes(
# TwoQubitWeylDecomposition(
# global phase: π/4
# ┌────────────┐ ┌─────────┐ ┌────────────┐┌──────────┐┌─────────────┐
# q_0: ┤ Rz(3.1389) ├───┤ Ry(π/2) ├──────────────┤0 ├┤ Ry(-π/2) ├┤ Rz(-1.5681) ├────────────
# └┬──────────┬┘┌──┴─────────┴──┐┌─────────┐│ Rxx(-π/2) │├─────────┬┘└┬────────────┤┌──────────┐
# q_1: ─┤ Rz(-π/2) ├─┤ Ry(0.0026466) ├┤ Rz(π/2) ├┤1 ├┤ Rz(π/2) ├──┤ Ry(1.5734) ├┤ Rz(-π/2) ├
# └──────────┘ └───────────────┘└─────────┘└────────────┘└─────────┘ └────────────┘└──────────┘
# )
b'k05VTVBZAQB2AHsnZGVzY3InOiAnPGMxNicsICdmb3J0cmFuX29yZGVyJzogRmFsc2UsICdzaGFw'
b'ZSc6ICg0LCA0KSwgfSAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAgICAg'
b'ICAgICAgICAgICAgIAoAAAAAAADwPwAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA'
b'AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA'
b'AAAAAAAAAAAAAAAAAAAAAAAA8D8AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA'
b'AAAAAAAAAAAA8D8AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA'
b'8D8AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA',
requested_fidelity=0.999999999,
calculated_fidelity=1.0,
actual_fidelity=1.0,
abc=(0.7853981633974483, 0.0, 0.0)),
'a': 0.7853981633974483,
'b': 0,
'c': 0,
'K1l': array([[ 9.99999781e-01-3.92003950e-17j, -7.84266235e-17-6.61646830e-04j],
[-7.84785659e-17-6.61646830e-04j, 9.99999781e-01+3.93042108e-17j]]),
'K1r': array([[ 0.5-0.5j, 0.5-0.5j],
[-0.5-0.5j, 0.5+0.5j]]),
'K2l': array([[ 7.07574481e-01+4.04613999e-20j, -4.05149776e-20-7.06638771e-01j],
[ 4.05149776e-20-7.06638771e-01j, 7.07574481e-01-4.04613999e-20j]]),
'K2r': array([[ 0.70710678+0.j, -0.70710678+0.j],
[ 0.70710678+0.j, 0.70710678+0.j]]),
'global_phase': 0.7853981633974487,
'unitary_matrix': array([[1.+0.j, 0.+0.j, 0.+0.j, 0.+0.j],
[0.+0.j, 0.+0.j, 0.+0.j, 1.+0.j],
[0.+0.j, 0.+0.j, 1.+0.j, 0.+0.j],
[0.+0.j, 1.+0.j, 0.+0.j, 0.+0.j]]),
'requested_fidelity': None,
'_is_flipped_from_original': False,
'calculated_fidelity': 1.0}This method contains a bit of randomness, which might be the source of the divergence, and can be also a good moment to review it (given the comment): https://github.com/Qiskit/qiskit-terra/blob/fe8ab43f1218d6997458fb3bf4011045f40932b0/qiskit/quantum_info/synthesis/two_qubit_decompose.py#L174-L175
Looking forward to know which result is the given code on your machine with the qiskit version used by me. Thanks in advance
levbishop
1ucian0
Environment
What is happening?
Running two supposedly equivalent variants of the same circuit leads to surprisingly discordant results. The divergence is very clear, it is reproducible on my machine, but not yet on another laptop (with Ubuntu 20.04, python 3.8, same conda packages).
How can we reproduce the issue?
Unfortunately this issue seems very much tied with some specific characteristic of my machine, since it doesn't happen with my other laptop. But it is yet unclear to me what is the specific problem. Anyway, on my machine I run the following setup and it always produce the failure:
Program A:
Program B: (equivalent because I add a circuit and its inverse).
Output:
Output:
Execute the program B:
What should happen?
The second program B, should create the same output of the first one: # OUTPUT: {'00000': 1024}
Any suggestions?
I tried two other setups other than the failing one on the remote machine.
SETUP A: (FAIL) as describe above + entire environment yaml below.
STEUP B: (NO FAIL) local system (aka laptop) with Ubuntu 20.04, python 3.8 and the same conda packages (except for few that were not available anymore, not sure why, marked below), Conda 4.11.0.
SETUP C: (NO FAIL) Dockerfile with Ubutnu 18.04.6 LTS, Miniconda3 with conda version 4.11.0, same conda packages, (except for few that were not available anymore, not sure why, marked below).
Hint: the error must be in the transpiler or qasm exporter, because when using the produced qasm file (visible above) it leads to failure (ALSO IN SETUP B and C)
I attach here the list of dump created via
conda env export > environment.yml: (Note that six pip packages were not installed in the second laptop, because they led to pip exceptions, they are marked with# LEADING TO PIP EXCEPTIONin the list).I am really looking forward to your feedback since I have no clue on why this happen and more importantly how to reproduce it on another machine, which is crucial for all of us to understand this bug.