An extended RZZ pass for unbounded parameters #2072

yaelbh · 2024-12-04T14:11:33Z

Closes #2032.
The transpiler pass of #2043 is extended to accommodate unbounded parameters, including parameter expressions.
Many thanks to Liran Shirizly for the idea.
The tests are currently failing - I'm investigating - opening in draft mode.

yaelbh · 2024-12-05T18:23:02Z

All the bugs have been fixed, it's not in draft mode anymore but ready for review.

yaelbh · 2024-12-06T05:11:41Z

Is there a risk that the pass will run again and again, forever?
Converting back to draft until this question is clarified.

wshanks

This is very clever, Yael and Liran.

I think the concern about running forever is coming from the use of the modified variable in a loop? That variable is just used to determine whether to reuse the original data or new data and not to restart the loop, so I think it is okay.

I have a couple concerns about this PR, but they may not be issues:

Should we be concerned about the blowing up in size of the input circuit?

The simple circuit with a single rzz with angle p goes from having a data attribute like:

[CircuitInstruction(operation=Instruction(name='rzz', num_qubits=2, num_clbits=0, params=[Parameter(p)]), qubits=(Qubit(QuantumRegister(2, 'q'), 0), Qubit(QuantumRegister(2, 'q'), 1)), clbits=())]

To one like:

[CircuitInstruction(operation=Instruction(name='global_phase', num_qubits=0, num_clbits=0, params=[ParameterExpression(1.5707963267949*(1 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5)))*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949)**2 + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5))**2 + 1.5707963267949*(sign(-p + 12.5663706143592*floor(0.0795774715459477*p) + 9.42477796076938) + 1)*((sign(p - 12.5663706143592*floor(0.0795774715459477*p) - 3.14159265358979) + 1)*sign(p - 12.5663706143592*floor(0.0795774715459477*p) - 3.14159265358979)**2/2 - sign(p - 12.5663706143592*floor(0.0795774715459477*p) - 3.14159265358979)**2 + 1)*sign(-p + 12.5663706143592*floor(0.0795774715459477*p) + 9.42477796076938)**2 - 0.785398163397448*(sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949) + 1)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2 - 0.785398163397448*(sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979) + 1)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2)]), qubits=(), clbits=()), CircuitInstruction(operation=Instruction(name='rz', num_qubits=1, num_clbits=0, params=[ParameterExpression(1.5707963267949*(sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949) + 1)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2 + 1.5707963267949*(sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979) + 1)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2)]), qubits=(Qubit(QuantumRegister(2, 'q'), 0),), clbits=()), CircuitInstruction(operation=Instruction(name='rx', num_qubits=1, num_clbits=0, params=[ParameterExpression(1.5707963267949*(1 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5)))*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949)**2 + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5))**2 + 1.5707963267949*(sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979) + 1)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2)]), qubits=(Qubit(QuantumRegister(2, 'q'), 0),), clbits=()), CircuitInstruction(operation=Instruction(name='rz', num_qubits=1, num_clbits=0, params=[ParameterExpression(1.5707963267949*(sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949) + 1)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2 + 1.5707963267949*(sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979) + 1)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2)]), qubits=(Qubit(QuantumRegister(2, 'q'), 1),), clbits=()), CircuitInstruction(operation=Instruction(name='rzz', num_qubits=2, num_clbits=0, params=[ParameterExpression(-(1 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5)))*(p - 6.28318530717959*floor(0.159154943091895*p + 0.5))*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949)**2 + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5))**2/2 + (p - 6.28318530717959*floor(0.159154943091895*p + 0.5))*(sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949) + 1)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5)) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5))**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5))**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949)**2/2 + (sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949) + 1)*(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2/2 + (sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979) + 1)*(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2/2)]), qubits=(Qubit(QuantumRegister(2, 'q'), 0), Qubit(QuantumRegister(2, 'q'), 1)), clbits=()), CircuitInstruction(operation=Instruction(name='rx', num_qubits=1, num_clbits=0, params=[ParameterExpression(1.5707963267949*(1 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5)))*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) + 1.5707963267949)**2 + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5))**2 + 1.5707963267949*(sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979) + 1)*((sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949) + 1)*sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2/2 - sign(p - 6.28318530717959*floor(0.159154943091895*p + 0.5) - 1.5707963267949)**2 + 1)*sign(-p + 6.28318530717959*floor(0.159154943091895*p + 0.5) + 3.14159265358979)**2)]), qubits=(Qubit(QuantumRegister(2, 'q'), 0),), clbits=())]

Perhaps in the weighing of pros and cons, this large expression is better than not being able to use parametrized circuits with rzz or needing to wrap parameter values in some external method.
Do we need to be careful about adding a link between the rzz gate and the rx gate? In practice, they were added at the same time and so rx is always present when rzz is. The first version of the pass just added an assumption of there being an XGate though, not RXGate. Perhaps it is fine the way it is. Or should there be a check that the backend supports RXGate as well to be safe?
I wonder if the parameterized and numerical code paths should be more aligned? Perhaps not, things are readable the way they are. The parameterized path uses a single function adding on parts as it goes while the numerical one has four separate functions. The parameterized version doesn't have the option of splitting up like the numerical one does.

One minor thing -- this is not related to this PR, but while looking at it I noticed the circuit drawings in the quad method docstrings are not aligned properly. Perhaps that could be incidentally fixed here as well.

yaelbh · 2024-12-08T10:39:16Z

For (2), I expect other translation passes to translate rx (or X in the numeric case) to basis gates, if they are not already basis gates. This is related to my infinite loop concern. I still have to learn how the pass manager works, but I imagine that the passes run and rerun until a fixed point. If this is the case, suppose that rx is not a basis gate. A pass modifies the DAG circuit to use basis gates instead, and, since the DAG circuit has been modified, our rzz pass runs again. Which generates new rx gates, and so on. Note that this issue does not apply to the numeric case, where only the first execution of the rzz pass modifies the DAG circuit.

For (1), I'm more concerned than you, to the point that I estimate that this will PR eventually will not be able to merge, because of the circuit size. This is certainly something that must be checked.

yaelbh · 2024-12-08T12:55:14Z

So the pass managers are just lists of passes that get to run sequentially, not involving loops and fixed points. I suggest at minimum to check if rx a basis gate, and skip the pass otherwise. We can also, in this case, create an alternative circuit without rx gates, but I find it unnecessary, since at least for now rx is present in all the backends that support fractional gates.

wshanks · 2024-12-09T16:04:43Z

qiskit_ibm_runtime/transpiler/passes/basis/fold_rzz_angle.py

@@ -46,7 +46,7 @@ class FoldRzzAngle(TransformationPass):
    with angle of arbitrary real numbers.
    """

-    def __init__(self, target: Optional[Union[Target, list[str]]] = None):
+    def __init__(self, target: Optional[Union[Target, List[str]]] = None):


If you prefer, you can use from __future__ import annotations at the top of the import statements and then the new style will work for type hints. Doing this causes the type hints to be converted into strings instead of being evaluated as objects (so it doesn't make the list[...] syntax actually work on Python 3.9; if you try to use it outside of an annotation, there will be an error).

I don't mind, do you have a preference?

We are inconsistent enough in this repo that I don't think it matters much in individual PRs like this, it would be more effective to have a PR that unifies across the repo and then we start being strict. My preference is to eventually settle on the new-style, in which case this would be Target | list[str] | None.

I don't have a strong preference though I do think it is nice to avoid the typing imports when possible. I mainly wanted to point the option out since I saw the commit message "old style typing required by CI".

ihincks · 2024-12-10T10:57:58Z

Should we be concerned about the blowing up in size of the input circuit?

Thanks for writing this example out explicitly, @wshanks . I don't have any feedback at the moment, just a question and some comments.

First of all, I haven't yet looked at the implementation or any math notes, but based on the gates in the printout, I assume this is implementing something the following case logic (?):

θ in [0, pi/2]: rzz(θ)
θ in [pi/2, pi]: (z \kron z) . rzz(θ-pi/2)
θ in [-pi/2, 0]: (x \kron I) . rzz(-θ) . (x \kron I) + global phase
θ in [-pi, -pi/2]: (x \kron i) . (z \kron z) . rzz(-θ-pi/2) . (x \kron I) + global phase

If so, I guess it's not surprising that implementing case logic with parameter expressions end up rather complicated. If there were many thousands of these inside of an input, it would cause bottlenecks various places in the stack.

To me, one of the main sticking points is where and how the user opts-in to the two extra rx gates.

yaelbh · 2024-12-10T11:11:49Z

The logic is described in a clear manner at the doc strings of the methods quad1 - quad4. Note that the rzz angle must always be valid. The rx gates come from the fact that X is only sometimes present (so rx angle is either 0 or pi), the user is not involved.

yaelbh added 4 commits December 4, 2024 15:19

extended rzz pass for unbounded parameters - a very initial commit

8e5981c

black

70b65bb

fix

abb73f3

fixes

6b6fe74

yaelbh marked this pull request as draft December 4, 2024 15:15

yaelbh added 8 commits December 5, 2024 15:22

cosmetics

0ac423e

a fix in the test

264e123

bug fix

2726561

bug fix

0c04455

removed dubug prints

3ffc8cd

lint

2d5c78c

mypy

62c869e

Merge branch 'main' into liran

58161a1

yaelbh marked this pull request as ready for review December 5, 2024 18:15

Merge branch 'main' into liran

0265df8

yaelbh requested a review from wshanks December 5, 2024 18:23

yaelbh marked this pull request as draft December 6, 2024 05:11

wshanks reviewed Dec 6, 2024

View reviewed changes

yaelbh added 6 commits December 9, 2024 11:09

don't convert to circuits with unsupported instructions

917055d

Merge branch 'tmp2' into liran

dbd87f9

aligned circuit drawings

b55ecd0

old style typing required by CI

4f7b1e1

black

afd0851

old style typing required by CI

5519c77

wshanks reviewed Dec 9, 2024

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

An extended RZZ pass for unbounded parameters #2072

An extended RZZ pass for unbounded parameters #2072

yaelbh commented Dec 4, 2024

yaelbh commented Dec 5, 2024

yaelbh commented Dec 6, 2024

wshanks left a comment

yaelbh commented Dec 8, 2024

yaelbh commented Dec 8, 2024

wshanks Dec 9, 2024

yaelbh Dec 10, 2024

ihincks Dec 10, 2024

wshanks Dec 10, 2024

ihincks commented Dec 10, 2024 •

edited

Loading

yaelbh commented Dec 10, 2024

An extended RZZ pass for unbounded parameters #2072

Are you sure you want to change the base?

An extended RZZ pass for unbounded parameters #2072

Conversation

yaelbh commented Dec 4, 2024

yaelbh commented Dec 5, 2024

yaelbh commented Dec 6, 2024

wshanks left a comment

Choose a reason for hiding this comment

yaelbh commented Dec 8, 2024

yaelbh commented Dec 8, 2024

wshanks Dec 9, 2024

Choose a reason for hiding this comment

yaelbh Dec 10, 2024

Choose a reason for hiding this comment

ihincks Dec 10, 2024

Choose a reason for hiding this comment

wshanks Dec 10, 2024

Choose a reason for hiding this comment

ihincks commented Dec 10, 2024 • edited Loading

yaelbh commented Dec 10, 2024

ihincks commented Dec 10, 2024 •

edited

Loading