Fix AerBackend issues caused by upgrading BackendV2 #1995
Conversation
doichanj
added
Changelog: Bugfix
doichanj
changed the title
qiskit_aer/backends/aer_simulator.py
Outdated
| @@ -518,6 +518,7 @@ class AerSimulator(AerBackend): | |||
| "continue_loop", | |||
| "reset", | |||
| "switch_case", | |||
| "barrier", | |||
There was a problem hiding this comment.
Barrier typically isn't needed in a target because qiskit always assumes barrier is valid for a backend since it's primarily a compiler directive. Was there something specific that was failing
There was a problem hiding this comment.
I removed barrier and add barrier inside plugin.
Adding barrier is needed because barrier can not be found in any name mapping and plugin returns with unsupported operation
qiskit_aer/backends/name_mapping.py
Outdated
| "u1": U1Gate, | ||
| "u2": U2Gate, | ||
| "u3": U3Gate, | ||
| "u": UGate, | ||
| "p": PhaseGate, | ||
| "r": RGate, | ||
| "rx": RXGate, | ||
| "rxx": RXXGate, | ||
| "ry": RYGate, | ||
| "ryy": RYYGate, | ||
| "rz": RZGate, | ||
| "rzz": RZZGate, | ||
| "rzx": RZXGate, | ||
| "id": IGate, | ||
| "h": HGate, | ||
| "x": XGate, | ||
| "y": YGate, | ||
| "z": ZGate, | ||
| "s": SGate, | ||
| "sdg": SdgGate, | ||
| "sx": SXGate, | ||
| "sxdg": SXdgGate, | ||
| "t": TGate, | ||
| "tdg": TdgGate, | ||
| "swap": SwapGate, | ||
| "cx": CXGate, | ||
| "cy": CYGate, | ||
| "cz": CZGate, | ||
| "ccx": CCXGate, | ||
| "csx": CSXGate, | ||
| "cp": CPhaseGate, | ||
| "cu": CUGate, | ||
| "cu1": CU1Gate, |
There was a problem hiding this comment.
For the standard gates they're covered by default from https://github.com/Qiskit/qiskit/blob/main/qiskit/circuit/library/standard_gates/__init__.py#L49 the custom name mapping passed to convert_to_target() is additive, the code takes get_standard_gate_name_mapping()'s output and runs .update(NAME_MAPPING) on it. Was there a reason you had to add these here too? Also in general if you pass in a gate class instead of an instance like this those instructions will be treated as variadic instructions which doesn't seem correct for standard library gates.
There was a problem hiding this comment.
removed standard gates from NAME_MAPPING
| circuit_to_dag(block), ops | ||
| ) | ||
| if node.name in self.config.target: | ||
| if node.name not in ops: |
There was a problem hiding this comment.
Doesn't this have a linear search overhead. Could you make ops a set instead I think everything else will work fine if it's a set.
| from qiskit_aer.backends.name_mapping import NAME_MAPPING | ||
|
|
||
|
|
||
| class AerBackendRebuildGateSetsFromCircuit(TransformationPass): |
There was a problem hiding this comment.
I think this should actually be an AnalysisPass you're not actually transforming the circuit, but are updating the target instead.
There was a problem hiding this comment.
Because there are only translation and scheduling plugins, I used translation plugin to analyze before going to optimization pass
| for name in ops: | ||
| if name not in self.config.target: | ||
| if name != "measure": | ||
| self.config.target.add_instruction(NAME_MAPPING[name], name=name) |
There was a problem hiding this comment.
I'm not sure this will do the correct thing in the case of a noise model. If you have a noise model won't this remove all the connectivity constraints caused by that? (this also connects with my comment below, you'd need to update the coupling map for any optimization or scheduling plugins).
Co-authored-by: Matthew Treinish <mtreinish@kortar.org>
Co-authored-by: Matthew Treinish <mtreinish@kortar.org>
| if self._target is not None: | ||
| return self._target | ||
|
|
||
| return convert_to_target(self.configuration(), self.properties(), None, NAME_MAPPING) |
There was a problem hiding this comment.
This is just a minor comment. Previously, converted target was cached. With this change, it will not be cached. I think frequent access of target property will produce performance regression. Introducing caching may help performance (though setting method may need clear cache of target).
There was a problem hiding this comment.
because if target is cached and used for transpilation to multiple circuits, basis gates for the first circuit is reused that can be a problem if the second circuit has gate that the first one does not have
* add description if no description is provided, build coupling map if it is provided * move import line * fix target for simulator backend * format * remove unused import * use translation plugin to rebuild gate sets for simulator * rename plugin * rebuild of gate sets is eanbled only for opt level 0 and 1 * fix custom pass manager * fix pass_manager function * added ccx in NAME_MAPPING * added missed gates in NAME_MAPPING * added release note * add check if opnodes is None * add check config * decrease return * check opt level * fix searching ops in control flow blocks * Update qiskit_aer/backends/plugin/aer_backend_plugin.py Co-authored-by: Matthew Treinish <mtreinish@kortar.org> * Update qiskit_aer/backends/plugin/aer_backend_plugin.py Co-authored-by: Matthew Treinish <mtreinish@kortar.org> * refer review comments * remove unused import --------- Co-authored-by: Matthew Treinish <mtreinish@kortar.org>
* add skip Python 3.12 for GPU build (#1965) * Fix basis gates of Aer backends (#1976) * move reset and switch_case ops to custom istr * Fix reset in AerStatevector * add test case * format * fix installing built Aer in some test cases * Applying global phase multiplication to initialize operation (#1980) * Applying global phase to initialize operation * fix format * remove recursive, add omp * Release 0.13.1 * Revert too many deprecations in Estimator (#1990) * Revert too much deprecation * fix typo * fix tests * Change priority of method selection of noise simulation (#1989) * Avoid selecting stabilizer method when noise model contains rotational gates * remove checking noise opsets, change priority selecting density_matrix * format * modify test cases use auto method result may change by this PR * modify one more test case * Remove use of opflow in Estimator (#1996) * Update misspelling apply_gate method doc (#1998) Co-authored-by: Jun Doi <doichan@jp.ibm.com> * Add optimization_level=0 to transpiler for compiling dynamic circuits (#2000) ``id`` gate was removed by transpiler called from aer_compiler without optimization_level for dynamic circuits. This commits adds ``optimization_level=0`` to avoid removing id gates * fix ry gate for stabilizer (#2001) Co-authored-by: Hiroshi Horii <hhorii@users.noreply.github.com> * Directly use psutil to get total system memory (#2002) Currently Aer is using Qiskit's local_hardware_info() function which to determine the total amount of system memory which is used to compute the largest statevector the system can build. However, this function wasn't really intended to be used outside of Qiskit and also Qiskit is looking to remove the memory reporting (see: Qiskit/qiskit#11254). This commit just pivots to using psutil directly which is what qiskit is doing internally. * test build fix (#2004) * Reverse ordering to read out error in sampling measure (#2003) * reverse ordering of read out error in sampling measure * fix batch check --------- Co-authored-by: Hiroshi Horii <hhorii@users.noreply.github.com> * Fix extended stabilizer thread safety in apply_ops_parallel (#1993) * Extended stabilizer simulator no longer shares RngEngine amongst states when ops are applied in parallel * Added release note * Fixed ugly cast --------- Co-authored-by: Jun Doi <doichan@jp.ibm.com> * add note (#1992) Co-authored-by: Jun Doi <doichan@jp.ibm.com> * Fix AerBackend issues caused by upgrading BackendV2 (#1995) * add description if no description is provided, build coupling map if it is provided * move import line * fix target for simulator backend * format * remove unused import * use translation plugin to rebuild gate sets for simulator * rename plugin * rebuild of gate sets is eanbled only for opt level 0 and 1 * fix custom pass manager * fix pass_manager function * added ccx in NAME_MAPPING * added missed gates in NAME_MAPPING * added release note * add check if opnodes is None * add check config * decrease return * check opt level * fix searching ops in control flow blocks * Update qiskit_aer/backends/plugin/aer_backend_plugin.py Co-authored-by: Matthew Treinish <mtreinish@kortar.org> * Update qiskit_aer/backends/plugin/aer_backend_plugin.py Co-authored-by: Matthew Treinish <mtreinish@kortar.org> * refer review comments * remove unused import --------- Co-authored-by: Matthew Treinish <mtreinish@kortar.org> * add prelude --------- Co-authored-by: Ikko Hamamura <ikkoham@users.noreply.github.com> Co-authored-by: Elena Peña Tapia <57907331+ElePT@users.noreply.github.com> Co-authored-by: jon <70080228+notcruz@users.noreply.github.com> Co-authored-by: Hiroshi Horii <hhorii@users.noreply.github.com> Co-authored-by: Matthew Treinish <mtreinish@kortar.org> Co-authored-by: eliotheinrich <38039898+eliotheinrich@users.noreply.github.com>
Summary
This PR fixes BackendV2 implementations in Aer compatible for Qiskit 0.45
Details and comments
This PR fixes issues :
AerSimulator.coupling_mapis not properly generated #1988Also this PR fixes test failure in Qiskit
test.python.circuit.test_controlled_gateit requires fix in Qiskit Qiskit/qiskit#11172For #1987 if input backend does not have description, it fails in Qiskit, so this fix add description.
For #1988, Aer did not build coupling map from option, so this fix check additional options and build coupling map
For #1982, added custom transpiler pass to rebuild basis gates including only gates appears in the input circuit, that prevent unnecessary change of gate. This is enabled when
optimization_level < 2