Test that gateset is preserved

[Misty-W]

Fixes #82 .

Adds a test to check if output circuits have gates outside of the gateset specified by the UCC BasisTranslator.

As expected, the test added here passes when CollectCliffords is commented out but otherwise fails on the random_clifford cases.

[jordandsullivan]

So do we know what the performance is if we convert the CliffordCollection back into 2 qubit gates? Is there still any advantage to using it?

[Misty-W]

So do we know what the performance is if we convert the CliffordCollection back into 2 qubit gates? Is there still any advantage to using it?

Testing that out now- making a modified CollectClffords pass looks to be the most straightforward way.

[jordandsullivan]

So it looks like the performance is not as good as Qiskit once again?

[Misty-W]

Yes, but that's using DefaultSynthesisClifford. I'm going to try some of the other synthesis passes to see if they improve the results.

[Misty-W]

We are still slightly worse than Qiskit in gate counts, but combining CollectCliffords and GreedySynthesisClifford does yield an improvement over UCC w/o that combination:

[jordandsullivan]

I see. I would like to understand how Qiskit intended for the Clifford object to be used, like is there a way to just take the info it contains as classical information that we keep track of, like a Clifford frame? Best, Jordan Sullivan

…

On Thu, Nov 14, 2024, 1:18 PM Misty Wahl ***@***.***> wrote: We are still slightly worse than Qiskit in gate counts, but combining CollectCliffords and GreedySynthesisClifford does yield an improvement over UCC w/o that combination: image.png (view on web) <https://github.com/user-attachments/assets/e593fd0f-8b31-41a4-85d5-6ae6a3023550> image.png (view on web) <https://github.com/user-attachments/assets/2c7780d4-f1b7-4ebd-bcb6-bad1803de434> — Reply to this email directly, view it on GitHub <#84 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADYYB5Y4ISZHPNQQQOJUDMT2AUHTDAVCNFSM6AAAAABRXQJK6OVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDINZXGQZDQMBUGE> . You are receiving this because you commented.Message ID: ***@***.***>

[Misty-W]

I see. I would like to understand how Qiskit intended for the Clifford object to be used, like is there a way to just take the info it contains as classical information that we keep track of, like a Clifford frame?

That's how I originally understood the purpose of the Clifford objects, but I need to dig further into the docs to confirm if such a framework is available. Meanwhile, I've found the synthesis (e.g. Clifford synthesis) passes work hand-in-hand with the collection functions, possibly indicating the collection functions being primarily for facilitating synthesis rather than for a hybrid workflow with classical simulation. The Qiskit compilation presentation at IBM QDC also showed collection and synthesis passes working in a similar fashion.

As an alternative, we could consider additional optimizations described in the paper on which greedy Clifford synthesis is based (https://arxiv.org/abs/2105.02291) that are not implemented in the current Qiskit version.

[Misty-W]

Using the workaround suggested in Qiskit/qiskit#11659 (comment) to force consolidation in ConsolidateBlocks, such that UnitarySynthesis works more like the Symbolic Peephole Optimization proposed in https://arxiv.org/pdf/2105.02291.

So, once again UCC reigns, albeit with a narrower margin and a caveat: we allow 1q and 2q Clifford gates (not Clifford objects or gates of >2 qubits), in addition to the default target basis.

[jordandsullivan]

So should we maybe add a check for whether the target gateset allows Cliffords (X,Y,Z,H) to decide whether to run this set of passes?

[Misty-W]

So should we maybe add a check for whether the target gateset allows Cliffords (X,Y,Z,H) to decide whether to run this set of passes?

I'm thinking an allow_Clifford_in_ouptut option that defaults to True.

[jordandsullivan]

I would prefer not to add another flag, since we already have a way to specify target gateset.

[Misty-W]

As a follow-up to this morning's meeting, I've removed the new passes including CollectCliffords from default and just kept the added test.

I opened PR #90 for adding back the new passes in a way that preserves the natural gateset.

[natestemen]

Two code clean up recommendations to help simplify the test, but looks like it's doing the right thing!

[natestemen]

Nice work! Important test to have.