[WIP] Use VF2 to find a partial layout for seeding a SabreLayout trial #10169
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 builds on the VF2PartialLayout pass which was an experiment available as an external plugin here: https://github.com/mtreinish/vf2_partial_layout That pass used the vf2 algorithm in rustworkx to find the deepest partial interaction graph of a circuit which is isomorphic with the coupling graph and uses that mapping to apply an initial layout. The issue with the performance of that pass was the selection of the qubits outside the partial interaction graph. Selecting the mapping for those qubits is similar to the same heuristic layout that SabreLayout is trying to solve, just for a subset of qubits. In VF2PartialLayout a simple nearest neighbor based approach was used for selecting qubits from the coupling graph for any virtual qubits outside the partial layout. In practice this ended up performing worse than SabreLayout. To address the shortcomings of that pass this commit combines the partial layout selection from that external plugin with SabreLayout. The sabre layout algorithm starts by randomly selecting a layout and then progressively working forwards and backwards across the circuit and swap mapping it to find the permutation caused by inserted swaps. Those permutations are then used to modify the random layout and eventual an initial layout that minimizes the number of swaps needed is selected. With this commit instead of using a completely random layout for all the initial guesses this starts a single trial with the partial layout found in the same way as VF2PartialLayout. Then the remaining qubits are selected at random and the Sabrelayout algorithm is run in the same manner as before. This hopefully should improve the quality of the results because we're starting from a partial layout that doesn't require swaps for those qubits. A similar (almost identical approach) was tried in Qiskit#9174 except instead of seeding a single trial with the partial layout it used the partial layout for all the the trials. In that case the results were not generally better and the results were mixed. At the time my guess was that using the partial layout constrained the search space too much and was inducing more swaps to be needed. However, looking at the details in issue Qiskit#10160 this adapts Qiskit#9174 to see if doing the partial layout in a more limited manner has any impact there.
mtreinish
force-pushed
the
vf2-partial-sabre
branch
from
b414a76
to
ca4c3bd
Compare
This commit drops the vf2 layout scoring from the sabre layout partial layout code. This code was just taken from the standalone vf2 passes (VF2Layout and VF2PostLayout) which tries multiple isomorphic subgraphs and figures out which layout has the best error characteristics. However, in the case of using a partial layout for sabre this is just wasted CPU cycles because we're not finalizing the layout and when we run sabre the actual layout used will be different. So we just need to pick any isomorphic subgraph and use that as a partial starting point. This commit removes all the multiple subgraph and scoring logic to improve the runtime performance with this branch.
This commit tweaks the algorithm for the initial layout guess when using a partial layout to use the nearest neighbor qubits instead of just a random order. This should lead to better results because there is less back and forth needed to bring the qubits together.
Pull Request Test Coverage Report for Build 5719974347
💛 - Coveralls |
This commit increases the heuristic effort for sabre layout and routing. This is made through 2 changes, the first is the depth of the internal lookahead heuristic used in sabre swap has been increased from 20 to 72. This is just a stop-gap for a potential reworking of the lookahead heuristic. In local testing for larger backends and deeper circuits in this is showing better output results than the current default of 20 without any runtime impact. The other aspect is that the trial counts for each optimization level > 0 is increased 5x. This has a runtime cost, but it the performance of the rust sabre implementation is fast enough that even running the algorithm with more trials it is not a bottleneck for typical compilation. Related to Qiskit#10160
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 commit builds on the VF2PartialLayout pass which was an experiment available as an external plugin here:
https://github.com/mtreinish/vf2_partial_layout
That pass used the vf2 algorithm in rustworkx to find the deepest partial interaction graph of a circuit which is isomorphic with the coupling graph and uses that mapping to apply an initial layout. The issue with the performance of that pass was the selection of the qubits outside the partial interaction graph. Selecting the mapping for those qubits is similar to the same heuristic layout that SabreLayout is trying to solve, just for a subset of qubits. In VF2PartialLayout a simple nearest neighbor based approach was used for selecting qubits from the coupling graph for any virtual qubits outside the partial layout. In practice this ended up performing worse than SabreLayout.
To address the shortcomings of that pass this commit combines the partial layout selection from that external plugin with SabreLayout. The sabre layout algorithm starts by randomly selecting a layout and then progressively working forwards and backwards across the circuit and swap mapping it to find the permutation caused by inserted swaps. Those permutations are then used to modify the random layout and eventual an initial layout that minimizes the number of swaps needed is selected. With this commit instead of using a completely random layout for all the initial guesses this starts a single trial with the partial layout found in the same way as VF2PartialLayout. Then the remaining qubits are selected at random and the Sabrelayout algorithm is run in the same manner as before. This hopefully should improve the quality of the results because we're starting from a partial layout that doesn't require swaps for those qubits.
A similar (almost identical approach) was tried in #9174 except instead of seeding a single trial with the partial layout it used the partial layout for all the the trials. In that case the results were not generally better and the results were mixed. At the time my guess was that using the partial layout constrained the search space too much and was inducing more swaps to be needed. However, looking at the details in issue #10160 this adapts #9174 to see if doing the partial layout in a more limited manner has any impact there.
Details and comments