Crosstalk adaptive gate scheduling pass #3532
Conversation
|
One of the builds fails because it doesn't find Z3. |
I think its also needed in the REQUIREMENTS list in setup.py. |
There was a problem hiding this comment.
Thanks Prakash. This looks pretty good and almost mergeable I think.
I made a few comments inline.
Can you add some release notes too? Instructions here.
There was a problem hiding this comment.
Right now if we're dependent on using z3-solver for this pass we'll need to make it optional. Adding z3-solver to requirements list would require osx users to have a C++ compiler installed because there are no precompiled wheels published for osx to pypi: Z3Prover/z3#2800 which is something we don't currently impose on any linux, windows, or osx users. Before we can move forward with this we should make z3 an optional dependency and not require it always be installed.
|
Thanks for making the change so quickly, the update lgtm. |
Co-Authored-By: Kevin Krsulich <kevin@krsulich.net>
There was a problem hiding this comment.
Ok this looks good to me now.
@kdk @lcapelluto can you check whether your comments are addressed?
| raise ImportError('z3-solver is required to use CrosstalkAdaptiveSchedule') | ||
| self.backend_prop = backend_prop | ||
| self.crosstalk_prop = crosstalk_prop | ||
| self.gate_id = {} |
There was a problem hiding this comment.
Through inspection, and because I also know the backend methods you used return time in s, it looks like you're using nanoseconds as the units for dur and time. Furthermore, the values are truncated at nanoseconds. If another developer looks at this code (or you, months/years from now), would that be apparent? Can you think of a way to make it obvious?
There was a problem hiding this comment.
Added a comment.
|
Thank you all for the comments and for merging the PR! |
* Implementation of crosstalk adaptive scheduling * Updated constructor params * fixed lint issues * fix test case bugs * fixed lint issues * updated setup.py requirements list with z3 * address review comments * address review comments * fixed parse_backend_properties * Update qiskit/transpiler/passes/crosstalk_adaptive_schedule.py Co-Authored-By: Kevin Krsulich <kevin@krsulich.net> * computing descendants and ancestors once per gate * fixed comments Co-authored-by: Kevin Krsulich <kevin@krsulich.net> Co-authored-by: Ali Javadi-Abhari <ajavadia@users.noreply.github.com>
This PR implements crosstalk adaptive gate scheduling.
This method aims to reducing crosstalk by serializing gate pairs which are prone to high crosstalk noise. However, naive serialization is bad because it incurs decoherence errors. So we use an optimization approach which takes an input circuit, backend crosstalk characterization data and backend properties to find an optimal instruction schedule that maximizes fidelity by considering both crosstalk and decoherence.
This is work done in collaboration with David McKay and Ali Javadi Abhari at IBM and Margaret Martonosi at Princeton. The paper is accepted for publication at ASPLOS 2020. arXiv link: https://arxiv.org/abs/2001.02826
I have added the tests to cover my changes.
I have updated the documentation accordingly.
I have read the CONTRIBUTING document.