Benchmarking SABRE transpilation

[geexie]

I'm currently experimenting with SABRE in Qibo. I saw a PR with SABRE optimisation with benchmarking results in it. May I know what benchmark was used to make the plot above? Do you have some data about circuit parameters as well like number of extra gates, etc? Which gate set do you use? Do you transpile for some specific architecture or Cliffird+T dataset?

I'm also wondering what transpilation pipeline setup is used for best SABRE performance? I use

    hw_gates = [gates.GPI2, gates.RZ, gates.Z, gates.CZ]
    natives = NativeGates(0).from_gatelist(hw_gates)
    custom_passes2 = [Preprocessing(), Unroller(native_gates=natives), Sabre(), Unroller(native_gates=natives)]
    return Passes(connectivity=nx.Graph(connectivity), passes=custom_passes2)

where connectivity map is one for IQM20

Thank you it advance!

[alecandido]

Hi @geexie, thanks for the interest.

That PR was authored by a temporary collaborator, which is inactive in Qibo since some time.
You may try to tag or reach him directly, but otherwise all the information we have is in that PR itself.

In any case, if you'd like to propose an alternative benchmark (possibly, more reproducible), it would be much appreciated.
Moreover, there is always room for contribution to Qibo transpiler. In case you are also interested in that, just let us know ;)
(feel free to write here, or also in private)

[geexie]

@alecandido, I see. I feel like the benchmark is here. It's not bad and moreover it's used in other transpiration papers. I just not sure about the Qibo transpiler configuration and connectivity map is used. @csookim any information would be appreciated.

I might be interested to contribute to transpiler later on.