-
Notifications
You must be signed in to change notification settings - Fork 2.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Stable] Improve 1q decomposition pass with multiple matching basis (#5431) #5655
[Stable] Improve 1q decomposition pass with multiple matching basis (#5431) #5655
Conversation
* Improve 1q decomposition pass with multiple matching basis This commit makes an improvement to the Optimize1qGatesDecomposition pass to improve the output from it in two scenarios. The first scenario is if there is an over complete basis where there are multiple choices for decomposition basis. Previously in such scenarios only the first matching subset would be used. This improves that by finding runs in all matching decomposition basis to simplify any runs in those gate sets. The second improvement made here is that the decomposition is only used if it results in a decrease in depth. There were scenarios where the general decomposition of a run would result in more gates than the input to the OneQubitEulerDecomposer (typically 3 gates vs 2 input). In those cases we shouldn't use the decomposition and instead rely on the original run because we're adding gates which is the opposit of the expected behavior. * Add DAGCircuit method to get 1q op node runs This commit adds a new DAGCircuit method, collect_1q_runs, which behaves like collect_runs(), but instead of collecting runs with gates on a name filter it looks at the number of qubits and returns runs of any op node that operate on a single qubit. This will be used by the optimize 1q decomposition pass to collect arbitrary single qubit gate runs and decompose them over the basis sets. * Switch to use all 1q runs instead of those just matching basis * Use retworkx.collect_runs for collect_1q_runs * Move parameterized gate split to collect_1q_runs * Reverse loops so basis is the inner loop and runs the outer * Ensure collect_1q_runs doesn't collect measurements * Only create temp circuit once per run * Move lone identity optimization into outer loop * Update qiskit/transpiler/passes/optimization/optimize_1q_decomposition.py * Fix lint * Use qc._append instead of qc.append Conflicts/Changes: For backport this makes 1 key change, the collect_1q_runs method is rewritten to be pure python instead of leveraging the retworkx collect_runs() method which was added in a newer version of retworkx than the minimum allowed version. This also partially pulls in the changes made in Qiskit#5570, but only some were backportable since the others relied on attributes that don't exist in gate objects in 0.16.x. (cherry picked from commit 5a1768f)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks good. Does this need its own release note?
@kdk Ok, I've updated this per your comments. As for the release notes I'll defer to you on it. I didn't write one for the master commit because I felt it was an internal implementation detail and kind of hard to describe the difference in a user facing perspective besides something like "update the 1q optimization decompose pass so it does a better job for an overcomplete basis" which didn't seem necessarily helpful as a fix note. |
The collect_1q_runs method was rewritten in python as part of the backport of a larger bugfix in the Optimize1qGatesDecomposition transpiler pass in Qiskit#5655. However, it is not part of the public api in 0.16.x and was only backported to fix the bug in Optimize1qGatesDecomposition. Since the version included on the master branch will be subtlely different in return type (after Qiskit#5685) Since we don't want to advertise the method until it's included in the public api this commit prepends '_' making the method name _collect_1q_runs to indicate it is internal only as part of the 0.16.2 release.
* Prepare 0.16.2 release This commit prepares for the 0.16.2 release by bumping the version strings to reflect the new version. This is just a bugfix release that several fixes for bugs introduced in 0.16.1. * Mark DAGCircuit collect_1q_runs private The collect_1q_runs method was rewritten in python as part of the backport of a larger bugfix in the Optimize1qGatesDecomposition transpiler pass in #5655. However, it is not part of the public api in 0.16.x and was only backported to fix the bug in Optimize1qGatesDecomposition. Since the version included on the master branch will be subtlely different in return type (after #5685) Since we don't want to advertise the method until it's included in the public api this commit prepends '_' making the method name _collect_1q_runs to indicate it is internal only as part of the 0.16.2 release. * Cleanup release notes
This commit fixes a bug introduced in the backport PR Qiskit#5655. On the current master branch the bugfix that was being backported relies on a feature in newer versions of retworkx. Since we can't bump the minimum retworkx version required on a stable branch that functionality was rewritten in the dagcircuit class directly. However, that implementation included a typo which is causing non-gates and atomic instructions to be included in collected runs which is not correct behavior and causes errors to be raised. This commit corrects the bug in the backport and adds a test to ensure the regression is not present. This is a direct commit to the stable branch and not a backport since this bug is not present on master and was only introduced through a bad backport. Fixes Qiskit#5736
This commit fixes a bug introduced in the backport PR #5655. On the current master branch the bugfix that was being backported relies on a feature in newer versions of retworkx. Since we can't bump the minimum retworkx version required on a stable branch that functionality was rewritten in the dagcircuit class directly. However, that implementation included a typo which is causing non-gates and atomic instructions to be included in collected runs which is not correct behavior and causes errors to be raised. This commit corrects the bug in the backport and adds a test to ensure the regression is not present. This is a direct commit to the stable branch and not a backport since this bug is not present on master and was only introduced through a bad backport. Fixes #5736
Summary
This commit makes an improvement to the Optimize1qGatesDecomposition
pass to improve the output from it in two scenarios. The first scenario
is if there is an over complete basis where there are multiple choices
for decomposition basis. Previously in such scenarios only the first
matching subset would be used. This improves that by finding runs in all
matching decomposition basis to simplify any runs in those gate sets.
The second improvement made here is that the decomposition is only used
if it results in a decrease in depth. There were scenarios where the
general decomposition of a run would result in more gates than the
input to the OneQubitEulerDecomposer (typically 3 gates vs 2 input). In
those cases we shouldn't use the decomposition and instead rely on the
original run because we're adding gates which is the opposit of the
expected behavior.
This commit adds a new DAGCircuit method, collect_1q_runs, which behaves
like collect_runs(), but instead of collecting runs with gates on a name
filter it looks at the number of qubits and returns runs of any op node
that operate on a single qubit. This will be used by the optimize 1q
decomposition pass to collect arbitrary single qubit gate runs and
decompose them over the basis sets.
Switch to use all 1q runs instead of those just matching basis
Move parameterized gate split to collect_1q_runs
Reverse loops so basis is the inner loop and runs the outer
Ensure collect_1q_runs doesn't collect measurements
Only create temp circuit once per run
Move lone identity optimization into outer loop
Update qiskit/transpiler/passes/optimization/optimize_1q_decomposition.py
Fix lint
Use qc._append instead of qc.append
Details and comments
Backported from #5431 and #5570
Conflicts/Changes:
For backport this makes 1 key change, the collect_1q_runs method is
rewritten to be pure python instead of leveraging the retworkx
collect_runs() method which was added in a newer version of retworkx
than the minimum allowed version. This also partially pulls in the
changes made in #5570, but only some were backportable since the others
relied on attributes that don't exist in gate objects in 0.16.x.
(cherry picked from commit 5a1768f)