Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Oxidize Commutation Analysis (#12995)
* init * up * lint * . * up * before cache * with cache * correct * cleaned up * lint reno * Update Cargo.lock * . * up * . * revert op * . * . * . * . * Delete Cargo.lock * . * corrected string comparison * removed Operator class from operation.rs * . * Apply suggestions from code review Co-authored-by: Raynel Sanchez <[email protected]> * comments from code review * Port DAGCircuit to Rust This commit migrates the entirety of the `DAGCircuit` class to Rust. It fully replaces the Python version of the class. The primary advantage of this migration is moving from a Python space rustworkx directed graph representation to a Rust space petgraph (the upstream library for rustworkx) directed graph. Moving the graph data structure to rust enables us to directly interact with the DAG directly from transpiler passes in Rust in the future. This will enable a significant speed-up in those transpiler passes. Additionally, this should also improve the memory footprint as the DAGCircuit no longer stores `DAGNode` instances, and instead stores a lighter enum NodeType, which simply contains a `PackedInstruction` or the wire objects directly. Internally, the new Rust-based `DAGCircuit` uses a `petgraph::StableGraph` with node weights of type `NodeType` and edge weights of type `Wire`. The NodeType enum contains variants for `QubitIn`, `QubitOut`, `ClbitIn`, `ClbitOut`, and `Operation`, which should save us from all of the `isinstance` checking previously needed when working with `DAGNode` Python instances. The `Wire` enum contains variants `Qubit`, `Clbit`, and `Var`. As the full Qiskit data model is not rust-native at this point while all the class code in the `DAGCircuit` exists in Rust now, there are still sections that rely on Python or actively run Python code via Rust to function. These typically involve anything that uses `condition`, control flow, classical vars, calibrations, bit/register manipulation, etc. In the future as we either migrate this functionality to Rust or deprecate and remove it this can be updated in place to avoid the use of Python. API access from Python-space remains in terms of `DAGNode` instances to maintain API compatibility with the Python implementation. However, internally, we convert to and deal in terms of NodeType. When the user requests a particular node via lookup or iteration, we inflate an ephemeral `DAGNode` based on the internal `NodeType` and give them that. This is very similar to what was done in #10827 when porting CircuitData to Rust. As part of this porting there are a few small differences to keep in mind with the new Rust implementation of DAGCircuit. The first is that the topological ordering is slightly different with the new DAGCircuit. Previously, the Python version of `DAGCircuit` using a lexicographical topological sort key which was basically `"0,1,0,2"` where the first `0,1` are qargs on qubit indices `0,1` for nodes and `0,2` are cargs on clbit indices `0,2`. However, the sort key has now changed to be `(&[Qubit(0), Qubit(1)], &[Clbit(0), Clbit(2)])` in rust in this case which for the most part should behave identically, but there are some edge cases that will appear where the sort order is different. It will always be a valid topological ordering as the lexicographical key is used as a tie breaker when generating a topological sort. But if you're relaying on the exact same sort order there will be differences after this PR. The second is that a lot of undocumented functionality in the DAGCircuit which previously worked because of Python's implicit support for interacting with data structures is no longer functional. For example, previously the `DAGCircuit.qubits` list could be set directly (as the circuit visualizers previously did), but this was never documented as supported (and would corrupt the DAGCircuit). Any functionality like this we'd have to explicit include in the Rust implementation and as they were not included in the documented public API this PR opted to remove the vast majority of this type of functionality. The last related thing might require future work to mitigate is that this PR breaks the linkage between `DAGNode` and the underlying `DAGCirucit` object. In the Python implementation the `DAGNode` objects were stored directly in the `DAGCircuit` and when an API method returned a `DAGNode` from the DAG it was a shared reference to the underlying object in the `DAGCircuit`. This meant if you mutated the `DAGNode` it would be reflected in the `DAGCircuit`. This was not always a sound usage of the API as the `DAGCircuit` was implicitly caching many attributes of the DAG and you should always be using the `DAGCircuit` API to mutate any nodes to prevent any corruption of the `DAGCircuit`. However, now as the underlying data store for nodes in the DAG are no longer the python space objects returned by `DAGCircuit` methods mutating a `DAGNode` will not make any change in the underlying `DAGCircuit`. This can come as quite the surprise at first, especially if you were relying on this side effect, even if it was unsound. It's also worth noting that 2 large pieces of functionality from rustworkx are included in this PR. These are the new files `rustworkx_core_vnext` and `dot_utils` which are rustworkx's VF2 implementation and its dot file generation. As there was not a rust interface exposed for this functionality from rustworkx-core there was no way to use these functions in rustworkx. Until these interfaces added to rustworkx-core in future releases we'll have to keep these local copies. The vf2 implementation is in progress in Qiskit/rustworkx#1235, but `dot_utils` might make sense to keep around longer term as it is slightly modified from the upstream rustworkx implementation to directly interface with `DAGCircuit` instead of a generic graph. Co-authored-by: Matthew Treinish <[email protected]> Co-authored-by: Raynel Sanchez <[email protected]> Co-authored-by: Elena Peña Tapia <[email protected]> Co-authored-by: Alexander Ivrii <[email protected]> Co-authored-by: Eli Arbel <[email protected]> Co-authored-by: John Lapeyre <[email protected]> Co-authored-by: Jake Lishman <[email protected]> * Update visual mpl circuit drawer references Right now there is a bug in the matplotlib circuit visualizer likely caused by the new `__eq__` implementation for `DAGOpNode` that didn't exist before were some gates are missing from the visualization. In the interest of unblocking this PR this commit updates the references for these cases temporarily until this issue is fixed. * Ensure DAGNode.sort_key is always a string Previously the sort_key attribute of the Python space DAGCircuit was incorrectly being set to `None` for rust generated node objects. This was done as for the default path the sort key is determined from the rust domain's representation of qubits and there is no analogous data in the Python object. However, this was indavertandly a breaking API change as sort_key is expected to always be a string. This commit adds a default string to use for all node types so that we always have a reasonable value that matches the typing of the class. A future step is likely to add back the `dag` kwarg to the node types and generate the string on the fly from the rust space data. * Make Python argument first in Param::eq and Param::is_close The standard function signature convention for functions that take a `py: Python` argument is to make the Python argument the first (or second after `&self`). The `Param::eq` and `Param::is_close` methods were not following this convention and had `py` as a later argument in the signature. This commit corrects the oversight. * Fix merge conflict with #12943 With the recent merge with main we pulled in #12943 which conflicted with the rust space API changes made in this PR branch. This commit updates the usage to conform with the new interface introduced in this PR. * Add release notes and test for invalid args on apply methods This commit adds several release notes to document this change. This includes a feature note to describe the high level change and the user facing benefit (mainly reduced memory consumption for DAGCircuits), two upgrade notes to document the differences with shared references caused by the new data structure, and a fix note documenting the fix for how qargs and cargs are handled on `.apply_operation_back()` and `.apply_operation_front()`. Along with the fix note a new unit test is added to serve as a regression test so that we don't accidentally allow adding cargs as qargs and vice versa in the future. * Restore `inplace` argument functionality for substitute_node() This commit restores the functionality of the `inplace` argument for `substitute_node()` and restores the tests validating the object identity when using the flag. This flag was originally excluded from the implementation because the Rust representation of the dag is not a shared reference with Python space and the flag doesn't really mean the same thing as there is always a second copy of the data for Python space now. The implementation here is cheating slighty as we're passed in the DAG node by reference it relies on that reference to update the input node at the same time we update the dag. Unlike the previous Python implementation where we were updating the node in place and the `inplace` argument was slightly faster because everything was done by reference. The rust space data is still a compressed copy of the data we return to Python so the `inplace` flag will be slightly more inefficient as we need to copy to update the Python space representation in addition to the rust version. * Revert needless dict() cast on metadata in dag_to_circuit() This commit removes an unecessary `dict()` cast on the `dag.metadata` when setting it on `QuantumCircuit.metadata` in `qiskit.converters.dag_to_circuit()`. This slipped in at some point during the development of this PR and it's not clear why, but it isn't needed so this removes it. * Add code comment for DAGOpNode.__eq__ parameter checking This commit adds a small inline code comment to make it clear why we skip parameter comparisons in DAGOpNode.__eq__ for python ops. It might not be clear why the value is hard coded to `true` in this case, as this check is done via Python so we don't need to duplicate it in rust space. * Raise a ValueError on DAGNode creation with invalid index This commit adds error checking to the DAGNode constructor to raise a PyValueError if the input index is not valid (any index < -1). Previously this would have panicked instead of raising a user catchable error. * Use macro argument to set python getter/setter name This commit updates the function names for `get__node_id` and `set__node_id` method to use a name that clippy is happy with and leverage the pyo3 macros to set the python space name correctly instead of using the implicit naming rules. * Remove Ord and PartialOrd derives from interner::Index The Ord and PartialOrd traits were originally added to the Index struct so they could be used for the sort key in lexicographical topological sorting. However, that approach was abandonded during the development of this PR and instead the expanded Qubit and Clbit indices were used instead. This left the ordering traits as unnecessary on Index and potentially misleading. This commit just opts to remove them as they're not needed anymore. * Fix missing nodes in matplotlib drawer. Previously, the change in equality for DAGNodes was causing nodes to clobber eachother in the matplotlib drawer's tracking data structures when used as keys to maps. To fix this, we ensure that all nodes have a unique ID across layers before constructing the matplotlib drawer. They actually of course _do_ in the original DAG, but we don't really care what the original IDs are, so we just make them up. Writing to _node_id on a DAGNode may seem odd, but it exists in the old Python API (prior to being ported to Rust) and doesn't actually mutate the DAG at all since DAGNodes are ephemeral. * Revert "Update visual mpl circuit drawer references" With the previous commit the bug in the matplotlib drawer causing the images to diverge should be fixed. This commit reverts the change to the reference images as there should be no difference now. This reverts commit 1e4e6f3. * Update visual mpl circuit drawer references for control flow circuits The earlier commit that "fixed" the drawers corrected the visualization to match expectations in most cases. However after restoring the references to what's on main several comparison tests with control flow in the circuit were still failing. The failure mode looks similar to the other cases, but across control flow blocks instead of at the circuit level. This commit temporarily updates the references of these to the state of what is generated currently to unblock CI. If/when we have a fix this commit can be reverted. * Apply suggestions from code review Co-authored-by: Raynel Sanchez <[email protected]> * code review * Fix edge cases in DAGOpNode.__eq__ This commit fixes a couple of edge cases in DAGOpNode.__eq__ method around the python interaction for the method. The first is that in the case where we had python object parameter types for the gates we weren't comparing them at all. This is fixed so we use python object equality for the params in this case. Then we were dropping the error handling in the case of using python for equality, this fixes it to return the error to users if the equality check fails. Finally a comment is added to explain the expected use case for `DAGOpNode.__eq__` and why parameter checking is more strict than elsewhere. * Remove Param::add() for global phase addition This commit removes the Param::add() method and instead adds a local private function to the `dag_circuit` module for doing global phase addition. Previously the `Param::add()` method was used solely for adding global phase in `DAGCircuit` and it took some shortcuts knowing that context. This made the method implementation ill suited as a general implementation. * More complete fix for matplotlib drawer. * Revert "Update visual mpl circuit drawer references for control flow circuits" This reverts commit 9a6f953. * Unify rayon versions in workspace * Remove unused _GLOBAL_NID. * Use global monotonic ID counter for ids in drawer The fundamental issue with matplotlib visualizations of control flow is that locally in the control flow block the nodes look the same but are stored in an outer circuit dictionary. If the gates are the same and on the same qubits and happen to have the same node id inside the different control flow blocks the drawer would think it's already drawn the node and skip it incorrectly. The previous fix for this didn't go far enough because it wasn't accounting for the recursive execution of the drawer for inner blocks (it also didn't account for LayerSpoolers of the same length). * Remove unused BitData iterator stuff. * Fully port Optimize1qGatesDecomposition to Rust This commit builds off of #12550 and the other data model in Rust infrastructure and migrates the Optimize1qGatesDecomposition pass to operate fully in Rust. The full path of the transpiler pass now never leaves rust until it has finished modifying the DAGCircuit. There is still some python interaction necessary to handle parts of the data model that are still in Python, mainly calibrations and parameter expressions (for global phase). But otherwise the entirety of the pass operates in rust now. This is just a first pass at the migration here, it moves the pass to be a single for loop in rust. The next steps here are to look at operating the pass in parallel. There is no data dependency between the optimizations being done by the pass so we should be able to the throughput of the pass by leveraging multithreading to handle each run in parallel. This commit does not attempt this though, because of the Python dependency and also the data structures around gates and the dag aren't really setup for multithreading yet and there likely will need to be some work to support that (this pass is a good candidate to work through the bugs on that). Part of #12208 * remove with_gil in favor of passing python tokens as params * Apply suggestions from code review Co-authored-by: Raynel Sanchez <[email protected]> * fmt * python serialization * deprecation * Update commutation_checker.py * heh * init * let Pytuple collect * lint * First set of comments - use Qubit/Clbit - more info on unsafe - update reno - use LazySet less - use OperationRef, avoid CircuitInstruction creation * Second part - clippy - no BigInt - more comments * Matrix speed & fix string sort -- could not use op.name() directly since sorted differently than Python, hence it's back to BigInt * have the Python implementation use Rust * lint & tools * remove unsafe blocks * One more try to avoid segfaulty windows -- if that doesn't work maybe revert the change the the Py CommChecker uses Rust * Original version Co-authored-by: Sebastian Brandhofer <[email protected]> * Sync with updated CommutationChecker todo: shouldn't make the qubits interner public * Debug: disable cache trying to figure out why the windows CI fails (after being unable to locally reproduce we're using CI with a reduced set of tests) * ... second try * Update crates/accelerate/src/commutation_checker.rs Co-authored-by: Raynel Sanchez <[email protected]> * Restore azure config * Remove unused import * Revert "Debug: disable cache" This reverts commit c564b80. * Don't overallocate cache We were allocating a the cache hashmap with a capacity for max cache size entries every time we instantiated a new CommutationChecker. The max cache size is 1 million. This meant we were allocating 162MB everytime CommutationChecker.__new__ was called, which includes each time we instantiate it manually (which happens once on import), the CommutationAnalysis pass gets instantiated (twice per preset pass manager created with level 2 or 3), or a commutation checker instance is pickle deserialized. This ends up causing a fairly large memory regression and is the source of the CI failures on windows. Co-authored-by: Jake Lishman <[email protected]> * Cleanup parameter key type to handle edge conditions better This commit cleans up the ParameterKey type and usage to make it handle edge conditions better. The first is that the type just doesn't do the right thing for NaN, -0, or the infinities. Canonicalization is added for hash on -0 and the only constructor of the newtype adds a runtime guard against NaN and inifinity (positive or negative) to avoid that issue. The approach only makes sense as the cache is really there to guard us against unnecessary re-computing when we reiterate over the circuit > 1 time and nothing has changed for gates. Otherwise comparing floats like done in this PR does would not be a sound or an effective approach. * Remove unnecessary cache hit rate tracking * Undo test assertion changes * Undo unrelated test changes * Undo pending deprecation and unify commutation classes This commit removes the pending deprecation decorator from the python class definition as the Python class just internally is using the rust implementation now. This also removes directly using the rust implementation for the standard commutation library global as using the python class is exactly the same now. We can revisit if there is anything we want to deprecate and remove in 2.0 in a follow up PR. Personally, I think the cache management methods are all we really want to remove as the cache should be an internal implementation detail and not part of the public interface. * Undo gha config changes * Make serialization explicit This commit makes the pickling of cache entries explicit. Previously it was relying on conversion traits which hid some of the complexity but this uses a pair of conversion functions instead. * Remove stray SAFETY comment * Remove ddt usage from the tests Now that the python commutation checker and the rust commutation checker are the same thing the ddt parameterization of the commutation checker tests was unecessary duplication. This commit removes the ddt usage to restore having a single run of all the tests. * Update release note * Fix CommutationChecker class import * Remove invalid test assertion for no longer public attribute * Ray's review comments Co-authored-by: Raynel Sanchez <[email protected]> * Handle ``atol/rtol``, more error propagation * update to latest changes in commchecker * fix merge conflict remnants * re-use expensive quantities such as the relative placement and the parameter hash * add missing header * gentler error handling * review comments & more docs * Use vec over IndexSet + clippy - vec<vec> is slightly faster than vec<indexset> - add custom types to satisfies clippy's complex type complaint - don't handle Clbit/Var * Simplify python class construction Since this PR was first written the split between the python side and rust side of the CommutationChecker class has changed so that there are no longer separate classes anymore. The implementations are unified and the python space class just wraps an inner rust object. However, the construction of the CommutationAnalysis pass was still written assuming there was the possibility to get either a rust or Python object. This commit fixes this and the type change on the `comm_checker` attribute by removing the unnecessary logic. --------- Co-authored-by: Raynel Sanchez <[email protected]> Co-authored-by: Kevin Hartman <[email protected]> Co-authored-by: Matthew Treinish <[email protected]> Co-authored-by: Elena Peña Tapia <[email protected]> Co-authored-by: Alexander Ivrii <[email protected]> Co-authored-by: Eli Arbel <[email protected]> Co-authored-by: John Lapeyre <[email protected]> Co-authored-by: Jake Lishman <[email protected]> Co-authored-by: Julien Gacon <[email protected]> Co-authored-by: Raynel Sanchez <[email protected]>
- Loading branch information
11 people
authored
Sep 4, 2024
1 parent
576efcf
commit 05f6429