routing_method='none' and Error pass

[1ucian0]

Fixes #4053

A new option routing_method='none' in the function qiskit.compiler.transpile.transpile was added to skip
any routing method. If the circuit does not fit coupling map, a qiskit.transpiler.exceptions.TranspilerError
exception will be raised.

The exception is raised via a new Error pass, as suggest in #3604 (comment)

This fixes #4756 as the problem in the issue was caused by the user unexpecting the routing map result. In that situation, the transpilation should be:

qc_joined = qiskit.transpile(qv_circs_joined, backend=backend, routing_method='none', optimization_level=0)

[kdk]

Looks good. Can you add tests for routing_method='none'?

[1ucian0]

Looks good. Can you add tests for routing_method='none'?

added in 0be9e9f

[ajavadia]

what's the problem exactly that routing_method='none' is trying to address? I had a look at the original issue (#4756) but couldn't really discern the issue. The circuit is not compatible with the backend coupling_map but the user wants to skip routing?

[ajavadia]

why is this a pass? it doesn't even operate on a circuit. An "ErrorPass" sounds weird. Trapping an error and returning it should happen outside the pass manager IMO.

[1ucian0]

we need a way to raise (in a broad sense) an error if dag is bad for some reason. We already have a way to see if a dag is bad, we have AnalysisPasses. Also, we want passes to "fix" those errors. A pattern like this:

pm.append(CheckProperty())
pm.append(TryToFix(), conditional=lambda ps: ps['is_bad'])
pm.append(CheckProperty())
pm.append(Error(), conditional=lambda ps: ps['is_bad'])

That pattern is similar to what happens here, when routing_method='none':

pm.append(CheckMap(coupling_map))
pm.append(Error(), condition=lambda property_set: not property_set['is_swap_mapped'])

Alternative, we could have ValidationPasss like suggested in #3604 . However, that will duplicate code with the AnalysisPass sometimes. For example, a ValidationMap pass (raises when the dag is not mapped) would have practically the same code as CheckMap.

To avoid the duplication, ValidationMap could require CheckMap. In that case, ValidationMap will also not operate on the circuit, but the property set. More relevantly, this approach will push the condition to check into the pass (instead of the passmanager condition). I think that's a fine solution, but it might break in a lot of validation passes with time.

[1ucian0]

Other possible examples for validation pass:

• Basis is not satisfiable CXDirection pass should not generate gates outside basis_gates #3972
• Circuit size/depth/gate count fits in backend Concurrent building of DAG, validation, and optimization when building quantum circuits  #1070
• Check gates after measure (backend dependant?) X Gate after measure impacts measurement on real backends.  #3289

[1ucian0]

what's the problem exactly that routing_method='none' is trying to address? I had a look at the original issue (#4756) but couldn't really discern the issue. The circuit is not compatible with the backend coupling_map but the user wants to skip routing?

@kdk linked #4053 about a similar discussion. I think there is value for the user to make sure no routing method is running, specially if the user runs a batch of circuits (likes in #4756) where a "bad" (as non compatible with the backend) circuit slips in.