implement swap update algorithm for adjacency constraint resolution when placing circuits on hardware #146
Conversation
…and QubitMapping data structures, and implement the swap update algorithm
…UpdateTransformer
| generate_decomposed_swap) | ||
| # First iteration adds a swap between Q0 and Q1. | ||
| # generate_decomposed_swap implements that swap operation as four | ||
| # sqrt-iswaps. |
There was a problem hiding this comment.
Instead of testing the implementation, we should test the effect. For instance, that the effect is equivalent to swapping Q[0] and Q[1] (via cirq.unitary or some other method).
There was a problem hiding this comment.
great idea.
Related question: how is the unitary matrix index determined for a circuit's qubits? For some of these tests, it would be cool to compare the logical circuit's unitary before and after adding swaps; they should be the same, but only after ignoring any ancillary qubits.
…nd add test coverage for previously untested public functions
|
Is this of general utility? Should this be upstreamed to cirq? |
FWIW, I had a university partner ask about routing code. I pointed them to cirq.contrib and here, and found it odd to point someone to quantum chess for this. |
Maybe we could try it out in Quantum chess and export it to cirq after some testing? |
That would be my preference as well. Happy to file a separate PR against cirq.contrib later but I'd prefer to do some evaluation first and iron out some of the kinks. |
Seems reasonable to me! |
|
|
||
| def generate_decomposed_swap( | ||
| q1: cirq.Qid, q2: cirq.Qid) -> Generator[cirq.Operation, None, None]: | ||
| """Generates a SWAP operation using sqrt-iswap gates.""" |
There was a problem hiding this comment.
Calling optimized_for_sycamore might be less efficient, but it's probably fine for now.
This is the second half of the algorithm described in this paper and is related to #127 and quantumlib/unitary#48
The swap updater takes an initial logical -> device qubit mapping and adds additional qubits and swap operations to ensure all gates operate on adjacent qubits. When combined with the initial placement implementation from #144 we'll get a complete circuit placement algorithm.