Add ElidePermutations pass to optimization level 3 #12111
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This commit adds a new transpiler pass ElideSwaps which is a logical optimization pass designed to run prior to layout or any other physical embedding steps in the transpiler pipeline. It traverses the DAG in topological order and when a swap gate is encountered it removes that gate and instead permutes the logical qubits for any subsequent gates in the DAG. This will eliminate any swaps in a circuit not caused by routing. Additionally, this pass is added to the preset pass manager for optimization level 3, we can consider adding it to other levels too if we think it makes sense (there is little overhead, and almost 0 if there are no swaps). One thing to note is that this pass does not recurse into control flow blocks at all, it treats them as black box operations. So if any control flow blocks contain swap gates those will not be optimized away. This change was made because mapping the permutation outside and inside any control flow block was larger in scope than what the intent for this pass was. That type of work is better suited for the routing passes which already have to reason about this.
The new ElideSwap pass is causing an output permutation just as a routing pass would. This information needs to be passed through to the output in case there is no layout or routing run. In those cases the information about the output permutation caused by the swap elision will be lost and doing layout aware operations like Operator.from_circuit() will not be able to reason about the permutation. This commit fixes this by inserting the original layout and qubit mapping into the property set along with the final layout. Then the base pass class and pass manager class are updated to use the original layout as the initial layout if one isn't set. In cases where we run layout and routing the output metadata from those passes will superscede these new metadata fields.
This commit fixes 2 issues in the execution of the new ElideSwaps pass as part of optimization level 3. First we were running it too late in the process both after high level synthesis (which isn't relavant yet, but will be soon when this is expanded to elide all permutations not just swaps) and also after reset diagonal before measurement. The second issue is that if the user is specifying to run with a manual layout set we should not run this pass, as it will interfere with the user intent.
There were 2 issues with the permutation tracking done in an earlier commit. First, it conflicted with the final_layout property set via routing (or internally by the routing done in the combined sabre layout). This was breaking conditional checks in the preset pass manager around embedding. To fix this a different property is used and set as the output final layout if no final layout is set. The second issue was the output layout object was not taking into account a set initial layout which will permute the qubits and cause the output to not be up to date. This is fixed by updating apply layout to apply the initial layout to the elision_final_layout in the property set.
This commit generalizes the pass from just considering swap gates to all permutations (via the PermutationGate class). This enables the pass to elide additional circuit permutations, not just the special case of a swap gate. The pass and module are renamed accordingly to ElidePermutations and elide_permutations respectively.
This commit fixes the recently added handling of the PermutationGate so that it correctly is mapping the qubits. The previous iteration of this messed up the mapping logic so it wasn't valid.
Co-authored-by: Jim Garrison <[email protected]>
Co-authored-by: Kevin Hartman <[email protected]>
jakelishman
reviewed
May 2, 2024
There was a problem hiding this comment.
If I understand the ordering correctly, the actual meat of this PR is commit 7a90b70. Assuming that's true, this mostly looks fine to me.
Comment on lines
411
to
415
| def generate_post_op_passmanager(): | ||
| """Generates the post-optimization pass manager.""" | ||
| out = PassManager() | ||
| out.append(FinalizeLayouts()) | ||
| return out |
There was a problem hiding this comment.
Highly unimportant, but just to point out that this can also be spelled return PassManager([FinalizeLayouts()]).
(No need to change it, just commenting in case it's shorter for you in the future.)
Comment on lines
109
to
111
|
|
||
| post_optimization = common.generate_post_op_passmanager() | ||
|
|
There was a problem hiding this comment.
I think we ought to include the layout finalisation in every optimisation level, even if by our presets it won't have any work to do. It just helps catch the case of somebody using (say) an O1 pass manager but injecting ElideSwaps into pre-op, which is a totally valid use of StagedPassManager.
Since we don't have final_permutation and its fixing of the update responsibilities to be local to the modifying pass, I think we need to treat FinalizeLayouts as a required part of the transpiler pipeline at all optimisation levels, or shift the responsibility of FinalizeLayouts into PassManager._passmanager_backend (#9523 (comment)).
There was a problem hiding this comment.
Since #9523 moved the responsibility of FinalizeLayouts into PassManager._passmanager_backend, now we don't need the post-op stage at all, rather than needing it everywhere.
There was a problem hiding this comment.
yes, this is much-much cleaner than having a separate FinalizeLayouts pass.
|
We don't need this PR anymore, should I close it? |
|
#9523 didn't add the pass to O3, so that part is still something we'd want to do separately. It's just that I think |
|
Ok, so the only thing that needs to be done (once ElidePermutations merges) is to replace OptimizeSwapsBeforeMeasure by ElidePermutations for level=3. |
mtreinish
reviewed
May 2, 2024
releasenotes/notes/add-elide-permutations-to-pipeline-077dad03bd55ab9c.yaml
Outdated
Show resolved
Hide resolved
mtreinish
previously approved these changes
May 2, 2024
mtreinish
reviewed
May 2, 2024
Comment on lines
+1767
to
+1772
| # an iswap gate is equivalent to (swap, CZ) up to single-qubit rotations. Normally, the swap gate | ||
| # in the circuit would cancel with the swap gate of the (swap, CZ), leaving a single CZ gate that | ||
| # can be realized via one ECR gate. However, with the introduction of ElideSwap, the swap gate | ||
| # cancellation can not occur anymore, thus requiring two ECR gates for the iswap gate. | ||
| self.assertEqual(res.count_ops()["ecr"], 2) | ||
| self.assertEqual(Operator(circuit), Operator.from_circuit(res)) |
There was a problem hiding this comment.
I guess this is the edge case of the elide permutations pass. It is a fun edge case and is unlikely to come up in practice I think.
…zation level 3, it does remove the SWAP gate. So we need to consider optimization_level=1 to assert that the extra pass does not remove SWAP gates
…qiskit-terra into elide-swaps-in-transpile
alexanderivrii
commented
May 2, 2024
Comment on lines
315
to
318
| spm = generate_preset_pass_manager( | ||
| optimization_level=3, initial_layout=list(range(2, -1, -1)), seed_transpiler=42 | ||
| optimization_level=1, initial_layout=list(range(2, -1, -1)), seed_transpiler=42 | ||
| ) | ||
| spm.layout += ElidePermutations() |
There was a problem hiding this comment.
This was the other new failing test. Now that the ElidePermutations already runs in level=3, it is able to remove those SWAP gates before getting to the extra pass in line 318.
mtreinish
approved these changes
May 2, 2024
sbrandhsn
added
the
Changelog: New Feature
ElePT
pushed a commit
to ElePT/qiskit
that referenced
this pull request
May 31, 2024
* Add ElideSwaps transpiler pass This commit adds a new transpiler pass ElideSwaps which is a logical optimization pass designed to run prior to layout or any other physical embedding steps in the transpiler pipeline. It traverses the DAG in topological order and when a swap gate is encountered it removes that gate and instead permutes the logical qubits for any subsequent gates in the DAG. This will eliminate any swaps in a circuit not caused by routing. Additionally, this pass is added to the preset pass manager for optimization level 3, we can consider adding it to other levels too if we think it makes sense (there is little overhead, and almost 0 if there are no swaps). One thing to note is that this pass does not recurse into control flow blocks at all, it treats them as black box operations. So if any control flow blocks contain swap gates those will not be optimized away. This change was made because mapping the permutation outside and inside any control flow block was larger in scope than what the intent for this pass was. That type of work is better suited for the routing passes which already have to reason about this. * Update tests with optimization level 3 * Pass final layout from ElideSwap to output The new ElideSwap pass is causing an output permutation just as a routing pass would. This information needs to be passed through to the output in case there is no layout or routing run. In those cases the information about the output permutation caused by the swap elision will be lost and doing layout aware operations like Operator.from_circuit() will not be able to reason about the permutation. This commit fixes this by inserting the original layout and qubit mapping into the property set along with the final layout. Then the base pass class and pass manager class are updated to use the original layout as the initial layout if one isn't set. In cases where we run layout and routing the output metadata from those passes will superscede these new metadata fields. * Move pass in opt level 3 earlier in stage and skip with explicit layout This commit fixes 2 issues in the execution of the new ElideSwaps pass as part of optimization level 3. First we were running it too late in the process both after high level synthesis (which isn't relavant yet, but will be soon when this is expanded to elide all permutations not just swaps) and also after reset diagonal before measurement. The second issue is that if the user is specifying to run with a manual layout set we should not run this pass, as it will interfere with the user intent. * Doc and copy paste fixes * Expand test coverage * Update permutation tracking There were 2 issues with the permutation tracking done in an earlier commit. First, it conflicted with the final_layout property set via routing (or internally by the routing done in the combined sabre layout). This was breaking conditional checks in the preset pass manager around embedding. To fix this a different property is used and set as the output final layout if no final layout is set. The second issue was the output layout object was not taking into account a set initial layout which will permute the qubits and cause the output to not be up to date. This is fixed by updating apply layout to apply the initial layout to the elision_final_layout in the property set. * Generalize pass to support PermutationGate too This commit generalizes the pass from just considering swap gates to all permutations (via the PermutationGate class). This enables the pass to elide additional circuit permutations, not just the special case of a swap gate. The pass and module are renamed accordingly to ElidePermutations and elide_permutations respectively. * Fix permutation handling This commit fixes the recently added handling of the PermutationGate so that it correctly is mapping the qubits. The previous iteration of this messed up the mapping logic so it wasn't valid. * Fix formatting * Fix final layout handling for no initial layout * Improve documentation and log a warning if run post layout * Fix final layout handling with no ElideSwaps being run * Fix docs build * Fix release note * Fix typo * Add test for routing and elide permutations * Apply suggestions from code review Co-authored-by: Jim Garrison <[email protected]> * Rename test file to test_elide_permutations.py * Apply suggestions from code review Co-authored-by: Kevin Hartman <[email protected]> * Fix test import after rebase * fixing failing test cases this should pass CI after merging Qiskit#12057 * addresses kehas comments - thx * Adding FinalyzeLayouts pass to pull the virtual circuit permutation from ElidePermutations to the final layout * formatting * partial rebase on top of 12057 + tests * also need test_operator for partial rebase * removing from transpiler flow for now; reworking elide tests * also adding permutation gate to the example * also temporarily reverting test_transpiler.py * update to release notes * minor fixes * adding ElidePermutations and FinalizeLayouts to level 3 * fixing tests * release notes * properly merging with main (now that ElidePermutations is merged) * and also deleting finalize_layouts * Update releasenotes/notes/add-elide-permutations-to-pipeline-077dad03bd55ab9c.yaml * updating code comment * Fixing failing test: now that ElidePermutations pass runs with optimization level 3, it does remove the SWAP gate. So we need to consider optimization_level=1 to assert that the extra pass does not remove SWAP gates --------- Co-authored-by: Matthew Treinish <[email protected]> Co-authored-by: Jim Garrison <[email protected]> Co-authored-by: Kevin Hartman <[email protected]> Co-authored-by: Sebastian Brandhofer <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary
This is a follow-up of #9523 that actually adds the
ElidePermutationsand the accompanyingFinalizeLayoutspasses to optimization level 3, and removes from there theOptimizeSwapBeforeMeasureas it's really just a special case ofElidePermutations(moreover,OptimizeSwapBeforeMeasuredoes not update the "final_layout" and breaks unitary equivalence between the original and the transpiled circuits).Details and comments
Should be rebased on top of #9523 (and in fact, separated from #9523 into a separate PR).