Make Sabre generic over analysis type#15799
Open
jakelishman wants to merge 7 commits intoQiskit:mainfrom
Open
Conversation
The documented behaviour of `SabreSwap` is that we assume that multi-qubit gates have already been decomposed. Previously, we silently ignored multi-qubit gates, most likely because this initial handling predated the introduction of the `directive` tracking in the Sabre interaction graph, so it was necessary to silently pass through barriers, which likely are multi-qubit. The new Sabre interaction graph does more work to categorise the types of interactions, with the intention of compressing interactions that can statically be known to have no impact on the routing, to avoid multiple Sabre trials from having to work through them. This means that Sabre now very clearly _sees_ a multi-qubit gate as it comes in, and can clearly distinguish it from a directive. The removed test case has always provided a nonsensical input to Sabre, as it includes 3q gates without decomposing them. The test was originally trying to assert that measures that appeared in the lexicographical output of the OpenQASM 2 exporter remained lexicographically after any inserted swaps by routing (see Qiskitgh-691 [1]). The swap mapper was not (in that case) violating the data-flow requirements, and the general case of forcing measurements into the final position was the inclusion of `BarrierBeforeFinalMeasurements` anyway, it was just so happening that the conversion from data-flow graph back to sequential data happened to go in a different order. [1]: d277521: Measurements before swap mapping
This allows `NLayout` to be indexed by `VirtualQubit` and `PhysicalQubit`, which is a shorter way of writing `VirtualQubit::to_phys` and `PhysicalQubit::to_virt`.
This commit separates out the three logical components of the Sabre routing tracking into three separate structs, where each struct groups objects that have the same mutation tendency. The previous Sabre state stored its problem description, internal tracking, and output tracking altogether in the same flat structure. Those three components have different tendencies to mutate: the problem description never mutates, the internal tracking frequently does, and the output tracking only occasionally does and has a lifetime validity tied to that of the problem description. Putting them together made it impossible to call methods that mutated the state while passing an object that borrowed from the problem description, such as when recursing into control-flow operations, because the borrow checker could not validate it. This applied interface pressure to inline more into the same method, which made code-reuse of separate concerns harder.
Sabre uses several objects that are logically maps from an index-like newtype (like `NodeIndex` or `PhysicalQubit`) to some value, and are implemented as fixed-slice `Vec`s for lookup efficiency. The newtype provides type safety while it's in use, but we have to cast it away to index, which makes it easy to index slices with the wrong object. This introduces a `VecMap` object, which provides a (minimal) slice-like interface, but indexes using the relevant newtype. The current implementation of Sabre does not use this _too_ much, but a refactoring of the layer structures will have them store one slice indexed by `PhysicalQubit` and one by `VirtualQubit`, which are trivially easy to get switched (a frequent mistake that is the base reason those new types were introduced in the first place).
The previous Sabre extended set was just the "next N" 2q gates topologically on from the front layer, where Qiskit reliably used `N = 20` ever since its introduction. For small-width circuits (as were common when the original Sabre paper was written, and when it was first implemented in Qiskit), this could mean the extended set was reliably several layers deep. This could also be the case for star-like circuits. For the wider circuits in use now, at the 100q order of magnitude, the 20-gate limit reliably means that denser circuits cannot have their entire next layer considered by the lookahead set. This commit modifies the lookahead heuristic to be based specifically on layers. This regularises much of the structure of the heuristic with respect to circuit and target topology; we reliably "look ahead" by the same "distance" as far as routing is concerned. It comes with the additional benefits: - we can use the same `Layer` structure for both the front layer and the lookahead layers, which reduces the amount of scoring code - the lookahead score of a swap can now affect at most two gates per layer, just like the front-layer scoring, and we can do this statically without loops - we no longer risk "biasing" the lookahead heuristic in either case of long chains of dependent gates (e.g. a gate that has 10 predecessors weights the score the same as a gate with only 1) or wide circuits (some qubits have their next layer counted in the score, but others don't because the extended set reached capacity). - applying a swap to the lookahead now has a time complexity that is constant per layer, regardless of the number of gates stored in it, whereas previously it was proportional to the number of gates stored (and the implementation in the parent of this commit is proportional to the number of qubits in the circuit). This change alone is mostly a set up, which enables further computational complexity improvements by modifying the lookahead layers in place after a gate routes, rather than rebuilding them from scratch, and subsequently only updating _swap scores_ based on routing changes, rather than recalculating all from scratch.
Ever since Sabre was first introduced to Qiskit (in [1]), the extended set has been recalculated from the front layer, every time the front layer is updated. Now that the extended set is tracked as a series of layers, it is easily possible to keep the layer structure updated in the same way as the front layer, by tracking separate points in the topological iteration through the interaction graph. To speed up synchronisation of the layers during updates, a new outer tracker, `Layers`, now means that a removal can jump immediately to the correct layer, without having to test multiples. This outer structure co-operates with the individual layers to effect faster lookups and removals; rather than each layer storing its own `IndexMap`, they instead store only a `Vec`, and the outer `Layers` tracks _where_ in the `Vec` any given gate is. On removal, the layer returns which, if any, gate has moved to occupy the otherwise empty slot (since removals are by swap, not shift). This avoids all hash lookups in the layer structures; everything is simple contiguous indexed access. With the tracking done elsewhere, and with `VecMap` providing type-safe indexing, each individual `Layer` now stores its contained gates in terms of `[VirtualQubit; 2]`; this minimises the amount of data that needs to be modified on a swap, at the cost that actions that need to associate actual gates with the current physical qubits (like iterating over active swaps) now need access to the `NLayout`, rather than the layout being (implicitly) tracked by each `Layer` via `apply_swap`. This is a performance improvement; as more overhead is reduced, the cost of the hash lookups in swap application was ever more visible. [1]: fab61d2: Sabre layout and routing transpiler passes (Qiskitgh-4537)
This lets callers of Sabre routing (`swap_map_trial`) inject the type of analysis they want performing, at compile time. This in turn lets Sabre layout avoid storing the full rebuild information during its forwards-backwards passes, when all it actually cares about is the swap counting and final layout (and really, it doesn't even care about the swap counting, but that's effectively free).
jakelishman
added
on hold
Collaborator
|
One or more of the following people are relevant to this code:
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This lets callers of Sabre routing (
swap_map_trial) inject the type of analysis they want performing, at compile time. This in turn lets Sabre layout avoid storing the full rebuild information during its forwards-backwards passes, when all it actually cares about is the swap counting and final layout (and really, it doesn't even care about the swap counting, but that's effectively free).Built on top of #14912.
On the HWB12 benchmark we recently merged to
main, this PR (on top of #14912) shows a modest (~2-3%) performance improvement inSabreLayout, which isn't a huge amount, but it isn't nothing.