optimization_level=1 always uses Trivial layout

[nonhermitian]

Information

• Qiskit Terra version: master
• Python version:
• Operating system:

What is the current behavior?

               ┌───┐                ░ ┌─┐            
q37_0: |0>─────┤ X ├────────────────░─┤M├────────────
          ┌───┐└─┬─┘┌───┐           ░ └╥┘┌─┐         
q37_1: |0>┤ H ├──■──┤ X ├───────────░──╫─┤M├─────────
          ├───┤     └─┬─┘┌───┐      ░  ║ └╥┘┌─┐      
q37_2: |0>┤ H ├───────■──┤ X ├──────░──╫──╫─┤M├──────
          ├───┤          └─┬─┘┌───┐ ░  ║  ║ └╥┘┌─┐   
q37_3: |0>┤ H ├────────────■──┤ X ├─░──╫──╫──╫─┤M├───
          ├───┤               └─┬─┘ ░  ║  ║  ║ └╥┘┌─┐
q37_4: |0>┤ H ├─────────────────■───░──╫──╫──╫──╫─┤M├
          └───┘                     ░  ║  ║  ║  ║ └╥┘
  c5_0: 0 ═════════════════════════════╩══╬══╬══╬══╬═
                                          ║  ║  ║  ║ 
  c5_1: 0 ════════════════════════════════╩══╬══╬══╬═
                                             ║  ║  ║ 
  c5_2: 0 ═══════════════════════════════════╩══╬══╬═
                                                ║  ║ 
  c5_3: 0 ══════════════════════════════════════╩══╬═
                                                   ║ 
  c5_4: 0 ═════════════════════════════════════════╩═

backend = provider.get_backend('ibmq_20_tokyo')
new_qc = transpile(qc, backend, optimization_level=1)

shows initial layout as the trivial layout, which is disconnected:

However part #2 in the source says _layout_check does better layout
if the circuit needs swaps,

# 2. Use a better layout on densely connected qubits, if circuit needs swaps

which is obviously the case here, but no other layouts
are actually selected (or imported).

Steps to reproduce the problem

What is the expected behavior?

Expected that the DenseLayout is used if there are swaps needed.

Suggested solutions

[1ucian0]

Confirmed.

qr = QuantumRegister(5)
cr = ClassicalRegister(5)
qc = QuantumCircuit(qr, cr)

qc.h(qr[1])
qc.h(qr[2])
qc.h(qr[3])
qc.h(qr[4])
qc.cx(qr[0], qr[1])
qc.cx(qr[1], qr[2])
qc.cx(qr[2], qr[3])
qc.cx(qr[3], qr[4])
qc.barrier(qr)
qc.measure(qr, cr)

backend = IBMQ.get_backend('ibmq_20_tokyo')
new_qc = transpile(qc, backend, optimization_level=1)
print('\n'.join([str((i, j)) for i, j in new_qc.layout._p2v.values() if i.name != 'ancilla']))

(QuantumRegister(5, 'q0'), 0)
(QuantumRegister(5, 'q0'), 1)
(QuantumRegister(5, 'q0'), 2)
(QuantumRegister(5, 'q0'), 3)
(QuantumRegister(5, 'q0'), 4)

Is also optimization_level=2 having a similar issue?

new_qc = transpile(qc, backend, optimization_level=2)
print('\n'.join([str((i, j)) for i, j in new_qc.layout._p2v.values() if i.name != 'ancilla']))

(QuantumRegister(5, 'q0'), 1)
(QuantumRegister(5, 'q0'), 4)
(QuantumRegister(5, 'q0'), 2)
(QuantumRegister(5, 'q0'), 3)
(QuantumRegister(5, 'q0'), 0)

[kdk]

Re-opened by #2975, will revisit post-0.9.

[1ucian0]

This can be fixed once "best of" approach is implemented (see #2969). On hold until then.

[1ucian0]

Under the assumption that DenseLayout is always better than TrivialLayout (is that true?), the PR #3657 is a simpler approach (if TriviaLayout does not return a perfect scoring, then it uses DenseLayout).