Unroll 3-or-moreQ pass

[1ucian0]

The current swappers do not work with 3Q gates. Therefore, it makes sense to decompose those gates before. This pass decomposes 3q gates recursively.

[jaygambetta]

This should not be just 3 qubit gates. We should have a pass that unrolls to 1 and 2 qubit gates. Why can't we call the unroll pass before swappers

[1ucian0]

The Unroll pass decomposes the full circuit to a basis set. That might create situations where some information get lost (for example, is this cx coming from a cz? ). This pass allows other passes to say: "I don't know how to deal with 3q, unroll just those ones."

[jaygambetta]

Then you can set it by the basis you want. If you are breaking up a cox you need to know what. For the mapper it will not mapper which gate basis you choose as long as it is single and two qubit.

[1ucian0]

Decompose and Unroll are complementary. Unroll('U') makes sure that you will only have 'U's at the end. Decompose('swap') will guarantee that you won't have 'swap's at the end. Both are necessary. Similarly to the former, Decompose3Q will keep everything as it is, but no 3Q at the end. That's what mapper needs. As you said, as long as the circuit has single or two qubits gate, the mapper is fine with it. There is no way to express that with the unroller.

[ajavadia]

It seems like what this pass should do is to unroll recursively down to a 1Q and 2Q basis. Decomposing just 3Q gates is limiting because what if the circuit has a 4Q or 5Q composite gate?
The pass can be written in a way that it decomposes until every gate is at maximum a two-body interaction. It is slightly different than the current unroller pass because no explicit basis is required.

[1ucian0]

I see your point there! It will no be testable until we don't have better support for custom gates. I can modify it to consider that and be read. Do you think the name DecomposeTo2QorLess is better?

[1ucian0]

After talking with @ajavadia and @jaygambetta, I see their point. The pass should be renamed to Unroll2Q since is recursive.

[1ucian0]

I renamed the pass and I change it to consider 3-or-moreQ gates, not just 3. Ready for review.

[ajavadia]

Thanks. This looks good. I updated the pass name to match the title of this PR, as I think that makes more sense. Good to merge.