I would like to ask how to set a particular layout during transpiling. I guess that the layout can be set by the initial_layout parameter in the transpiler. However, there are several options that may conflict, namely: layout_method, and optimisation optimization_level. I do not know which one suppress the other. I guess that optimization_level setting to 0 can enforce it. But on the other hand I still want to a bit of optimising anything else apart from the layout. I search out for the documentation but there seems to be not much talking about this. Any help is very appreciated.


Disclaimer: I might be plain wrong:) More experienced qiskit users/developers are encouraged comment!

I guess initially I was thinking about layouts the way are thinking right now. I assumed that the choice of initial_layout binds the logical qubits from my original circuit to the physical qubits and this binding remains intact during the computation. However, a better way to think about this is that logical qubits are constantly moving around. Although you should be able to say at each moment which physical qubit stores a given logical qubit this mapping can change dynamically. Moreover, it may be not possible to keep it constant because of the connectivity constraints or even if it is possible, it might not be efficient.

Let us consider a simple example, a CNOT gate

enter image description here

Assume that you can only apply two-quit gates between $q_0,q_1$ and $q_1,q_2$. Then this circuit could be made compatible with this connectivity by adding some SWAPs

enter image description here

Even if in the initial layout your control qubit was mapped to $q_0$ after the first SWAP it sits at $q_1$. Yes, we can then perform our CNOT and return the state of the control qubit back to $q_0$ (by the second SWAP, as shown on the image), but this might not be efficient. Suppose that you want to make two gates between $q_0$ and $q_2$, like here

enter image description here

If you insist that your initial layout be restored after each logical gate you can do it as follows (at the barrier the initial layout is restored)

enter image description here

However, this is obviously inefficient as the two subsequent SWAPs combine to an identity.

OK, so qiskit has a lot of clever algorithms to avoid this kind inefficiency and thy involve frequent rearrangements of logical qubits. Initial layout simply determines how do you start. If you still insist on keeping the layout fixed you should probably explain why and, in the case of constrained connectivity, in what sense?

p.s. A gotcha. Precisely because of the reasons I described qiskit's transpiler does not attempt to restore the initial layout at the end of the transpilation. If you measure your state in the end, it just maps the physical qubits to the classical registers according to the last layout. However, at some point I was interested in the unitary matrix of the circuit Transpilation on restricted topology does not yield an equivalent circuit in Qiskit , and without measurements I did not find a way to figure out the final layout. So in this sense the transpilation of a circuit does not always preserve the unitary matrix of the circuit.

  • 1
    $\begingroup$ The last statement is not quite correct. The transpiler does preserve unitarity, but only up to a permutation matrix. Figuring out the final layout has been on the wish list for some time and is straightforward to do if you just keep track of the swap permutations. After proposing this topic for many hackatoons and having no takers I did it myself here: github.com/Qiskit/qiskit-terra/pull/6827. $\endgroup$ Sep 15 at 9:00
  • $\begingroup$ Thank you for your comment. Having a manual layout is important when one wants to calibrate errors of the qubits. If the clibration works on some physical qubits and the actual circuit run on some different qubits, it makes then no sense. It is interesting to me that qiskit tutorial does not mention this at all; and it is not wrong because it is running on a simulator (without simulated constraints). But when one wants to do mitigation for an actual quantum computer, one has to take care of that (I think). $\endgroup$
    – soara
    Sep 15 at 10:27
  • $\begingroup$ @PaulNation Sure, fair point, the unitary is preserved up to a permutation. (Or is it? As far as I remember some transpiler passes can assume implicitly that that the input state is all zeros, and act accordingly. So for example a single $Z$ gate at the beginning of a circuit might get removed.) Is there a way to use your PR until it is merged in the official version? $\endgroup$ Sep 15 at 11:40
  • $\begingroup$ @soara, OK I think I understand your motivation. However, I guess that for the purposes of calibration it makes most sense to not change the circuit at all. Once you allow any change/transpilation, there may be no clear-cut difference between changing a layout and other transformations. In the end, layout change is just a bunch of SWAP gates. If you want to prohibit this, what transformations do you allow? In my last figure, is cancelling adjacent CNOTs allowed or not? $\endgroup$ Sep 15 at 11:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.