Describe extension to gate functionality

[ndebeaudrap]

Summary

This PR changes the open specification, consistent with the proposals in (the latter half of) Issue #323. This is submitted following discussions in the Circuit Families and Generics WG and in the TSC meetings, and collaboration with the Types and Casting WG.

Details and comments

We introduce a supplemental, more versatile syntax for defining gate subroutines, and a more versatile specification of what a gate body may consist of.

• This allows the classical parameters of a gate to have different types (the 'default' type of a gate is considered retrospectively to correspond to an angle).*
• It also allows the quantum operands of a gate to consist of qubit registers of a specific size (which are not considered to be subject to broadcasting).

To do this, the additional declaration syntax is expanded to permit the programmer to explicitly specify the types of all of the parameters and operands. Example gate invocations are provided to demonstrate how such gates may be invoked, and under what conditions the invocations are considered to be well-typed.

(* That the parameters of gate declarations are of type angle, is a tentative position reached by the Types and Casting WG. Discussion of this point should involve other members of that WG. Note that the more versatile declaration syntax described in this PR, allows direct support of gate subroutines which accept float parameters instead, for those cases where this would be desirable.)

The gate body is allowed to contain program logic in the form of for loops and if statements, which may involve the classical gate parameters (or expressions which use those parameters), so long as these can be determined at compile-time. It is also allowed to involve declaration of classical storage types, subject to the same constrants as def subroutines, albeit with the limitation that locally defined identifiers (and classical parameters) may not be re-assigned values after they are initialised.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
2 out of 3 committers have signed the CLA.

✅ ndebeaudrap
✅ hodgestar
❌ Jonathan Robert Niel de Beaudrap

---

Jonathan Robert Niel de Beaudrap seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[jwoehr]

@ndebeaudrap , there seems to be some automation fu blocking you signing the CLA ... can you please click the " add the email address used for this commit to your account " link and straighten out the angry machine?!

[ndebeaudrap]

It appears to me, novice Git user that I am, that I did not configure my email properly on my local system. I do not believe that the email message in the commit is one that I would be able to 'verify', and that this is the root of the problem.

I am attempting a fix, having configured my local system better. If this does not work, I would appreciate advice on the best way forward in producing a better formed PR along these lines, given the current state of it.

[ndebeaudrap]

(If — as I suspect — it would be expedient to simply remove this branch, please do so; I will try again once that has been done.)

[jwoehr]

(If — as I suspect — it would be expedient to simply remove this branch, please do so; I will try again once that has been done.)

Looks like it's fixed.

[stevenheidel]

Nit: which can be resolved - it's required not optional that resolution exists to a sequence of built-in gates.

[ndebeaudrap]

Noted and amended

[stevenheidel]

Nit: matching style from previous examples, there's no space after the gate name

[ndebeaudrap]

Noted and amended

[stevenheidel]

Hm, this breaks the design principle we had discussed where a compiler could choose to represent gates as unitary matrices in its implementation. To such a compiler, the gates:

   gate g : qubit a
   {
     U(0, 0, 0) a;
   } 

would be the same as

   gate g : qubit a
   {
     delay[20ns] a;
     U(0, 0, 0) a;
   } 

[blakejohnson]

That doesn't change the unitary meaning of the gate, though, does it? (assuming delay is equivalent to an identity matrix)

[blakejohnson]

I guess the alternatives would connect delay to some kind of drift operator (e.g. a static coupling term) or a Kraus operator representing amplitude dampling and/or dephasing. But, all of these break out of the model of gates having a 1-1 correspondence with unitaries.

[stevenheidel]

The point is that they look the same in one representation but they look different in another representation. Meaning a compiler storing gates as unitaries would think the gates interchangeable even though they aren't.

On the other hand that could already be accomplished today by adding an identity gate to a gate definition.

When a programmer puts a delay inside a gate definition, I'd guess that they really want a delay and would be unhappy if the compiler were to come up with a different implementation of the gate.

[ndebeaudrap]

Fair point; amended.

[blakejohnson]

^^ N.B. to other reviewers

[stevenheidel]

👍 I believe this is consistent with what we discussed at the TSC meeting.

[taalexander]

I have no major objections to this, other than raising the question of if this should be required for version 1.0?

[levbishop]

My concern with this is that by changing from treating the parameters as exact angles to considering them as inexact types, then we can no longer represent the mathematical equivalence between parameterized gates, which is the intention of the gate keyword. Example:

# from some common gate library:
gate g1(a, b) q {
    U(a*5, b/5, 0) q;
}
gate g2(a, b) q {
    U(a*3, b/3, 0) q;
}

# from hardware definition:
defcal g2(angle[13] a, angle[13] b) $0 {...}

# from user code:
angle[12] phi = pi/7;
angle[12] lam = pi/7;
g1(phi, lam) $0;

We'd like a compiler to be able to see this and infer the equivalence between g1 and g2 and rewrite g1(phi, lam) as g2(3*phi/5, 5*lam/3) which can then match with the defcal and the program can execute. The calculation of 3*phi/5 will be done at angle[12] precision which is what the author of the user code asked for and presumably they have determined this is sufficient precision for their algorithm, so that seems fine (maybe¹). When it goes into the defcal that result is cast to angle[13] because that is the best the hardware can support, which also seems fine and unavoidable.

If instead the include files have:

gate g1(angle[8] a, angle[8] b) q {
    U(a*5, b/5, 0) q;
}
gate g2(angle[8] a, angle[8] b) q {
    U(a*3, b/3, 0) q;
}

then the equivalence is unclear:

• Does the compiler search for an angle[8] x such that the rounded x*5 is equal to angle[8] y = angle[8](angle[12](pi/7))*3, and then cast back to angle[12]? What if there is no such angle[8] x?
• Does the compiler search for an angle[8] x such that the rounded x/5 is equal to angle[8] y = angle[8](angle[12](pi/7))/3, and then cast back to angle[12]? What if there is more than one such angle[8] x? (This one seems less of a problem - since we already have multiple ways of implementing the same unitary, eg trivially U(x, y, z) has equivalences mod 2pi, and all of U(0, x, -x) are identity: we trust the compiler to choose something sensible for these cases)
• Does the compiler just calculate 3*phi/5 in angle[12] precision like in the arbitrary-precision gate case? If so, then what is the purpose of the type annotation in the gate definition?
• On what basis is the precision angle[8] (or any other precision in gate definitions) to be chosen? It would be nice to write gate in a generic way that's independent of both the user algorithm and the hardware implementation. For example, I think we'd like to give a gate crz(theta) c, t in hardware/algortihm independent way as in stdgates.inc.

Footnotes

1. hmm. maybe assuming the precision of the user variable is appropriate for gate-rewriting is not ideal. What if the user algorithm is doing something like 4 rounds of iterative phase estimation and thus defines the phase as angle [4] phase_est and does the bit twiddling on the 4 bits of the angle to make that work. If the gates involving phase_est need to be rewritten for the hardware, the user probably still wants that rewriting done at higher accuracy than angle[4]. Maybe it makes more sense to do all of the rewriting at the precision of the defcal, angle[13] in the example, since that is the ultimate precision that goes to the hardware anyway? (if so, what if the hardware has different precision for phases and amplitudes...?) ↩

[levbishop]

Pulled the parameter typing question out into a separate issue #375.
General consensus from discussion at the last TSC meeting was that the remainder of this PR is in good shape and can. be ready to merge after we have some decision on how to proceed on #375.

[ndebeaudrap]

I believe, subject to the resolution of #375 (and in particular the presumption that OpenQASM2-style gate declarations take angles as parameters), this PR is ready to be merged. Do we concur?

[bettinaheim]

Looks good to me. I think this is just waiting for the updates from Lev, and then this should tie in nicely.

[blakejohnson]

I assume this will need to be rebased after #393 is merged, but I agree that this seems fine to proceed following that change.

[blakejohnson]

#393 is finally merged. Can we refactor this change set accordingly?

[blakejohnson]

@ndebeaudrap are you available to re-factor these changes on top of gates.rst as it sits now?

[hodgestar]

I've merged main into this branch and resolved conflicts. The text likely needs a final round of clean ups and reviews. I'm happy to do the initial round of clean ups but reviews and comments welcomed.

[blakejohnson]

Thank you @hodgestar for doing that work. I will add re-reading this to my TODO list.

[braised-babbage]

A common justification for distinguishing gates and subroutines is that a gate is a kind of mathematical specification that can be manipulated by a compiler. For a gate definition with no arguments this is a matrix with constant entries. For a gate definition with angle arguments the matrix entries may be symbolic expressions (already suggesting that a compiler may need to support symbolic algebra of some sort). When we further enrich the gate definition language with if statements and for loops which may involve gate arguments, the relationship between a gate definition and the corresponding matrix may be highly nontrivial.

It seems like this extension would require that a compiler handle certain gate applications by instantiating them as sub-circuits (i.e. substituting in arguments and performing compile-time evaluation of if statements and for loops, replacing the application with the result of this). At that point, why restrict ourselves to unitary operations? Many interesting sub-circuits involve measurements. OpenQASM subroutines do not have restrictions on whether or not expressions in them can be evaluated at compile time (assuming const arguments), which is part of the reason for this proposal.

[blakejohnson]

I think the intent here was still to be able to associate generalized gates with unitary matrices, but allowing convenient short hands. I suppose we have failed to do that, though, because it is trivial to construct examples where this is no longer true, e.g.:

gate classically_controlled_not(bit b): qubit q {
    if (b) { x q; }
}

I'm not sure if there are additional constraints we could introduce to restore this property? Or do we just give up on trying to keep it and allow gates to not always be associated with unitaries?

[meamy]

Just weighing in here since this popped up in my inbox --- I don't see the example above as disassociating gates from unitary matrices. Rather, a gate is a family of unitary matrices defined by the values of all classical parameters. In this sense it's morally no different (I believe) from the interpretation of openQASM 2.0 style gates, it's just allowing other types of parameters and a different type of computation if the compiler actually intended on generating the unitary given a set of parameters. This proposal was a long time ago however so my memory of what was proposed and the issues may be a bit fuzzy.

[blakejohnson]

I see your point, @meamy. The thing which is changed is that I no longer have a mathematical structure for the associated unitary which is fixed other than for substitution of a variable into some algebraic expressions. But, my example is definitely still unitary for any value of b.

[meamy]

Indeed, you could say it takes it from a matrix with symbolic expressions inside to a symbolic expression of a matrix. The question comes down to which is more useful in practice, and how that balances against the perspective of the programmer (if openQASM is intended to be a programming language rather than an IR) and other considerations like broadcasting.

[blakejohnson]

Prior to this change set, the discussion of controlled gates sat under a "quantum gate modifiers" section. I'd suggest preserving that organizational structure.

[blakejohnson]

Actually, this entire section of looks like it refers to a stale version of the "quantum gate modifiers" section. I believe this change should be reverted.

[braised-babbage]

Agreed. For example, the broadcasting discussion below is already discussed adequately elsewhere https://openqasm.com/language/gates.html#broadcasting.

[blakejohnson]

This change looks fine.

[blakejohnson]

This looks like the start of the changes we want to keep.

[blakejohnson]

The // comment isn't part of the gate declaration, no?

[blakejohnson]

I struggled to scan "for loop" as a compound noun, and so was confused by the parenthetical statement. Would it be clearer to just say "(Loop induction variables ...)"?

[blakejohnson]

This entire section can be replaced with a reference to the broadcasting section above.

[blakejohnson]

Having refreshed my memory on this topic by re-reading the PR, I am still broadly in favor of this change. However, I think @braised-babbage put his finger on the critical issue of whether this change breaks the correspondence between gates and unitary matrices. If so, we either need to accept that, or figure out other changes to the proposal.

[braised-babbage]

Following up on our discussion of this at the last TSC, there are a few conceptual pieces here.

1. The introduction of a typed parameter list for parametric gates (allowing classical parameters beyond angles).
2. The introduction of a typed quantum argument list for gates (allowing more compact and sophisticated notion of gate applications involving qubit registers).
3. An enrichment of what is allowed in the gate body (namely, if and for statements which involve only only const or gate parameter values in a way that can be resolved at compile time).

To a degree, these items can be considered separately.

For example, item (1) above can be useful when considering the interaction between OpenQASM's gate and defcal constructs. Whereas a gate formally represents a unitary operation, a defcal provides an operational description of how to perform such an operation. At present, a defcal cannot be parameterized by meaningful quantities (e.g. DAC scale factors, IQ classification thresholds, etc) that may otherwise be manipulated by a low-level OpenPulse program because these quantities are not of angle type. Without such parametrization, calibration definitions necessarily rely on a certain amount of global state. For this reason I find (1) to be of some interest, even if we do not expect that the corresponding gate definitions to have any mathematical dependence on these classical quantities.

While there is an interaction between (2) and (3), (e.g. if we have a qubit register we can loop over, this motivates supporting for statements in a gate body), we can still consider them separately. Fundamentally (3) is of interest to people defining gates, whereas (2) is also of interest to people invoking gates. Specifically considering (2), the current broadcasting mechanism can allow for compact expression e.g. of transversal operations, but is still relatively transparent: when you read a gate application, you can tell from the arguments what is going on. When we allow (2), we lose this property -- a reader of the program is ultimately going to have to dig in to the gate definition to make sense of it. This is the price of the added flexibility.

[blakejohnson]

I think we need to be more precise about what a "simple expression" is.

[braised-babbage]

Following up on today's TSC meeting (8/7/24), I created an issue #547 representing a kind of "item 0" for #346 (comment).