Fix potential infinite loop in SabreSwap #7970
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 adds a "release valve" mechanism to `SabreSwap`, so that it will always eventually make forwards progress and terminate the run. Before this commit, it was possible for certain pathological circuits (see `looping_circuit` in `test/python/transpiler/test_sabre_swap.py` for an example) to get stuck in a stable local minimum of the 'lookahead' or 'decay' heuristics (no matter the weightings). The release mechanism is done with very small per-loop overhead for the vastly more common good paths; we simply count how many iterations we have been through since we made progress, and the minimum weight in the coupling map to be overcome to make progress again. Once the number of iterations without progress exceeds some value, we backtrack (inefficiently, because this is the bad path) to the last time we made progress, and forcibly insert swaps to bring the nearest gate together. We then continue like normal. There are also a few minor optimisations in this commit that prevent recalculating sets that we already know; `extended_set` is fixed by knowledge of the `front_layer` if the DAG isn't (meaningfully) changed, so there's no need to recalculate it on each loop.
jakelishman
added
the
Changelog: Bugfix
|
Thank you for opening a new pull request. Before your PR can be merged it will first need to pass continuous integration tests and be reviewed. Sometimes the review process can be slow, so please be patient. While you're waiting, please feel free to review other open PRs. While only a subset of people are authorized to approve pull requests for merging, everyone is encouraged to review open pull requests. Doing reviews helps reduce the burden on the core team and helps make the project's code better for everyone. One or more of the the following people are requested to review this:
|
Pull Request Test Coverage Report for Build 2202565163
💛 - Coveralls |
ajavadia
reviewed
Apr 21, 2022
mtreinish
added
the
stable backport potential
mtreinish
reviewed
Apr 21, 2022
Co-authored-by: Matthew Treinish <[email protected]>
The `max_iterations_without_progress` value was originally exposed to users via the constructor. It was added because I thought I'd need a way to force the backtracking algorithm to occur to test the behaviour. When I wrote the tests, I ended up using Aer to find the keys, which is sufficiently fast that we can verify validity using the regular infinite loop circuit, and the argument is unnecessary. It was highly unlikely to actually be useful in real applications, so it does not need to be added.
mtreinish
approved these changes
Apr 21, 2022
Comment on lines
+452
to
+457
| split = len(path) // 2 | ||
| forwards, backwards = path[1:split], reversed(path[split:-1]) | ||
| for swap in forwards: | ||
| yield v_start, layout._p2v[swap] | ||
| for swap in backwards: | ||
| yield v_goal, layout._p2v[swap] |
There was a problem hiding this comment.
Just a comment for the record this hurt my head for a little while until I realized it's intended to work on virtual bits and that the layout is updated on each yield. So it's not the sequence of physical bits to add swaps on
mergify bot
pushed a commit
that referenced
this pull request
Apr 21, 2022
* Fix potential infinite loop in SabreSwap This adds a "release valve" mechanism to `SabreSwap`, so that it will always eventually make forwards progress and terminate the run. Before this commit, it was possible for certain pathological circuits (see `looping_circuit` in `test/python/transpiler/test_sabre_swap.py` for an example) to get stuck in a stable local minimum of the 'lookahead' or 'decay' heuristics (no matter the weightings). The release mechanism is done with very small per-loop overhead for the vastly more common good paths; we simply count how many iterations we have been through since we made progress, and the minimum weight in the coupling map to be overcome to make progress again. Once the number of iterations without progress exceeds some value, we backtrack (inefficiently, because this is the bad path) to the last time we made progress, and forcibly insert swaps to bring the nearest gate together. We then continue like normal. There are also a few minor optimisations in this commit that prevent recalculating sets that we already know; `extended_set` is fixed by knowledge of the `front_layer` if the DAG isn't (meaningfully) changed, so there's no need to recalculate it on each loop. * Reduce depth of greedy swap insertion * Add link in release note Co-authored-by: Matthew Treinish <[email protected]> * Remove extra keyword argument in SabreSwap constructor The `max_iterations_without_progress` value was originally exposed to users via the constructor. It was added because I thought I'd need a way to force the backtracking algorithm to occur to test the behaviour. When I wrote the tests, I ended up using Aer to find the keys, which is sufficiently fast that we can verify validity using the regular infinite loop circuit, and the argument is unnecessary. It was highly unlikely to actually be useful in real applications, so it does not need to be added. * Remove useless return from _add_greedy_swaps Co-authored-by: Matthew Treinish <[email protected]> (cherry picked from commit f414273)
mergify bot
added a commit
that referenced
this pull request
Apr 21, 2022
* Fix potential infinite loop in SabreSwap This adds a "release valve" mechanism to `SabreSwap`, so that it will always eventually make forwards progress and terminate the run. Before this commit, it was possible for certain pathological circuits (see `looping_circuit` in `test/python/transpiler/test_sabre_swap.py` for an example) to get stuck in a stable local minimum of the 'lookahead' or 'decay' heuristics (no matter the weightings). The release mechanism is done with very small per-loop overhead for the vastly more common good paths; we simply count how many iterations we have been through since we made progress, and the minimum weight in the coupling map to be overcome to make progress again. Once the number of iterations without progress exceeds some value, we backtrack (inefficiently, because this is the bad path) to the last time we made progress, and forcibly insert swaps to bring the nearest gate together. We then continue like normal. There are also a few minor optimisations in this commit that prevent recalculating sets that we already know; `extended_set` is fixed by knowledge of the `front_layer` if the DAG isn't (meaningfully) changed, so there's no need to recalculate it on each loop. * Reduce depth of greedy swap insertion * Add link in release note Co-authored-by: Matthew Treinish <[email protected]> * Remove extra keyword argument in SabreSwap constructor The `max_iterations_without_progress` value was originally exposed to users via the constructor. It was added because I thought I'd need a way to force the backtracking algorithm to occur to test the behaviour. When I wrote the tests, I ended up using Aer to find the keys, which is sufficiently fast that we can verify validity using the regular infinite loop circuit, and the argument is unnecessary. It was highly unlikely to actually be useful in real applications, so it does not need to be added. * Remove useless return from _add_greedy_swaps Co-authored-by: Matthew Treinish <[email protected]> (cherry picked from commit f414273) Co-authored-by: Jake Lishman <[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 adds a "release valve" mechanism to
SabreSwap, so that it willalways eventually make forwards progress and terminate the run. Before
this commit, it was possible for certain pathological circuits
(see
looping_circuitintest/python/transpiler/test_sabre_swap.pyfor an example) to get stuck in a stable local minimum of the
'lookahead' or 'decay' heuristics (no matter the weightings).
The release mechanism is done with very small per-loop overhead for the
vastly more common good paths; we simply count how many iterations we
have been through since we made progress, and the minimum weight in the
coupling map to be overcome to make progress again. Once the
number of iterations without progress exceeds some value, we backtrack
(inefficiently, because this is the bad path) to the last time we made
progress, and forcibly insert swaps to bring the nearest gate together.
We then continue like normal.
There are also a few minor optimisations in this commit that prevent
recalculating sets that we already know;
extended_setis fixed byknowledge of the
front_layerif the DAG isn't (meaningfully) changed,so there's no need to recalculate it on each loop.
Details and comments
Fix #7707.
This commit also knocks off some of the low-hanging fruit in the efficiency of the
SabreSwapimplementation, since I had to touch it anyway; we don't recalculate theextended_seton each iteration if no progress on the first layer has been made, because it shouldn't have changed. We just store it and move on. This is good for a 25-30% speedup inSabreSwapfor 20 qubits at a depth of 1024, and it passes some savings on to a generaltranspile(optimization_level=3), which is most notable for QV circuits.