Run logical optimization passes in init stage #11354
Conversation
This commit adds Optimize1qGatesDecomposition and InverseCancellation to the init stage for optimization levels 1, 2, and 3. The optimizations these passes run are completely valid on abstract circuits and do not require target qubits to be selected to run (`Optimize1qGatesDecomposition` is noise aware and will perform target aware optimizations if one is provided, but if no target information is reported it will operate correctly for any circuit)
There was a problem hiding this comment.
In principle I totally agree with doing this. Personally I think that abstract optimisation is / ought to be a sufficiently important step that I'd go as far as wanting a new named stage for it between init and layout, but I can live with it being init at least initially.
Are you sure this doesn't warrant a release note? It doesn't technically affect the public API, but I definitely think it's a change worth drawing attention to - if nothing else, it's a somewhat modification of the obligations of the init stage.
| @@ -88,6 +88,44 @@ def pass_manager(self, pass_manager_config, optimization_level=None) -> PassMana | |||
| pass_manager_config.unitary_synthesis_plugin_config, | |||
| pass_manager_config.hls_config, | |||
| ) | |||
| elif optimization_level in {1, 2}: | |||
There was a problem hiding this comment.
Incredibly minor: there's no reason to use a set here - optimization_level in (1, 2) would technically be more efficient. Not important at all, though.
| else: | ||
| return TranspilerError(f"Invalid optimization level {optimization_level}") |
There was a problem hiding this comment.
Completely unrelated, but imo this should actually be a ValueError.
|
Let's add a logical optimization stage. I think it's totally reasonable to break this out into a stage. FWIW, the description and expectations of the init stage is currently: which I felt included this. But I think you're right we probably should separate this out form EDIT: But maybe as a follow up stage because I forgot this will require adding another plugin hook point and extra documentation and stuff which is larger in scope than this change. |
In Qiskit#11354 we added more logical optimization passes to the preset pass managers for level 1, 2, and 3 which included the Optimize1qGatesDecomposition pass. However, this had the the unintended side effect of tending to normalize circuits to use parameterized gates. This becomes an issue for discrete basis targets as the basis translator doesn't always know how to work with those. To avoid issues this commit removes the pass from the init stage and just saves it for the optimization loop. If we want to do 1q gate simplification during init in the future we can develop a different pass that does it in a manner that wouldn't interfere with discrete basis targets like Optimize1qGatesDecomposition.
In #11354 we added more logical optimization passes to the preset pass managers for level 1, 2, and 3 which included the Optimize1qGatesDecomposition pass. However, this had the the unintended side effect of tending to normalize circuits to use parameterized gates. This becomes an issue for discrete basis targets as the basis translator doesn't always know how to work with those. To avoid issues this commit removes the pass from the init stage and just saves it for the optimization loop. If we want to do 1q gate simplification during init in the future we can develop a different pass that does it in a manner that wouldn't interfere with discrete basis targets like Optimize1qGatesDecomposition.
Summary
This commit adds Optimize1qGatesDecomposition and InverseCancellation to the init stage for optimization levels 1, 2, and 3. The optimizations these passes run are completely valid on abstract circuits and do not require target qubits to be selected to run
(
Optimize1qGatesDecompositionis noise aware and will perform target aware optimizations if one is provided, but if no target information is reported it will operate correctly for any circuit)Details and comments