qiskit 2.0.1__cp39-abi3-macosx_11_0_arm64.whl → 2.1.0rc1__cp39-abi3-macosx_11_0_arm64.whl

qiskit/transpiler/passes/scheduling/padding/context_aware_dynamical_decoupling.py

@@ -0,0 +1,876 @@
+# This code is part of Qiskit.
+#
+# (C) Copyright IBM 2024.
+#
+# This code is licensed under the Apache License, Version 2.0. You may
+# obtain a copy of this license in the LICENSE.txt file in the root directory
+# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
+#
+# Any modifications or derivative works of this code must retain this
+# copyright notice, and modified files need to carry a notice indicating
+# that they have been altered from the originals.
+
+"""Context-aware dynamical decoupling."""
+
+from __future__ import annotations
+from enum import Enum
+
+import itertools
+import logging
+
+from pprint import pformat
+from dataclasses import dataclass, field
+from collections import defaultdict
+import numpy as np
+import rustworkx as rx
+
+from qiskit.circuit import QuantumCircuit, Qubit, Reset, Gate
+from qiskit.circuit.library import CXGate, ECRGate, XGate
+from qiskit.dagcircuit import DAGCircuit
+from qiskit.converters import circuit_to_dag
+from qiskit.dagcircuit import DAGOpNode, DAGInNode
+from qiskit.transpiler.target import Target
+from qiskit.transpiler.basepasses import TransformationPass
+from qiskit.transpiler.coupling import CouplingMap
+from qiskit.transpiler.exceptions import TranspilerError
+
+from .pad_delay import PadDelay
+
+logger = logging.getLogger(__name__)
+
+
+class ContextAwareDynamicalDecoupling(TransformationPass):
+    """Implement an X-sequence dynamical decoupling considering the gate- and qubit-context.
+
+    This pass implements a context-aware dynamical decoupling (DD) [1], which ensures that
+
+        (1) simultaneously occurring DD sequences on device-adjacent qubits are mutually orthogonal, and
+        (2) DD sequences on spectator qubits of ECR and CX gates are orthogonal to the echo
+            pulses on the neighboring target/control qubits.
+
+    The mutually orthogonal DD sequences are currently Walsh-Hadamard sequences, consisting of only
+    X gates. In some cases it might therefore be beneficial to use :class:`.PadDynamicalDecoupling`
+    with more generic sequences, such as XY4.
+
+    This pass performs best if the two-qubit interactions have the same durations on the
+    device, as it allows to align delay sequences and take into account potential control and target
+    operations on neighboring qubits. However, it is still valid if this is not the case.
+
+    .. note::
+
+        If this pass is run within a pass manager (as in the example below), it will
+        automatically run :class:`.PadDelay` to allocate the delays. If instead it is run as
+        standalone (not inside a :class:`.PassManager`), the delays must already be inserted.
+
+
+    Example::
+
+        from qiskit.circuit import QuantumCircuit
+        from qiskit.circuit.library import QFTGate
+        from qiskit.transpiler import PassManager
+        from qiskit.transpiler.passes import ALAPScheduleAnalysis, ContextAwareDynamicalDecoupling
+        from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager
+        from qiskit_ibm_runtime.fake_provider import FakeSherbrooke
+
+        num_qubits = 10
+        circuit = QuantumCircuit(num_qubits)
+        circuit.append(QFTGate(num_qubits), circuit.qubits)
+        circuit.measure_all()
+
+        target = FakeSherbrooke().target
+
+        pm = generate_preset_pass_manager(optimization_level=2, target=target)
+        dd = PassManager([
+            ALAPScheduleAnalysis(target=target),
+            ContextAwareDynamicalDecoupling(target=target),
+        ])
+
+        transpiled = pm.run(circuit)
+        with_dd = dd.run(transpiled)
+
+        print(with_dd.draw(idle_wires=False))
+
+    References:
+
+        [1] A. Seif et al. (2024). Suppressing Correlated Noise in Quantum Computers via
+            Context-Aware Compiling, `arXiv:2403.06852 <https://arxiv.org/abs/2403.06852>`_.
+
+    """
+
+    def __init__(
+        self,
+        target: Target,
+        *,
+        min_duration: int | None = None,
+        skip_reset_qubits: bool = True,
+        skip_dd_threshold: float = 1.0,
+        pulse_alignment: int | None = None,
+        coloring_strategy: rx.ColoringStrategy = rx.ColoringStrategy.Saturation,
+    ) -> None:
+        """
+        Args:
+            target: The :class:`.Target` of the device to run the circuit.
+            min_duration: Minimal delay duration (in ``dt``) to insert a DD sequence. This
+                can be useful, e.g. if a big delay block would be interrupted and split into
+                smaller blocks due to a very short, adjacent delay. If ``None``, this is set
+                to be at least twice the difference of the longest/shortest CX or ECR gate.
+            skip_reset_qubits: Skip initial delays and delays after a reset.
+            skip_dd_threshold: Skip dynamical decoupling on an idle qubit, if the duration of
+                the decoupling sequence exceeds this fraction of the idle window. For example, to
+                skip a DD sequence if it would take up more than 95% of the idle time, set this
+                value to 0.95. A value of 1. means that the DD sequence is applied if it fits into
+                the window.
+            pulse_alignment: The hardware constraints (in ``dt``) for gate timing allocation.
+                If provided, the inserted gates will only be executed on integer multiples of
+                this value. This is usually provided on ``backend.configuration().timing_constraints``.
+                If ``None``, this is extracted from the ``target``.
+            coloring_strategy: The coloring strategy used for ``rx.greedy_graph_color``.
+                Defaults to a saturation strategy, which is optimal on bipartite graphs,
+                see Section 1.2.2.8 of [2].
+
+        References:
+
+            [2] A. Kosowski, and K. Manuszewski, Classical Coloring of Graphs, Graph Colorings,
+                2-19, 2004. ISBN 0-8218-3458-4.
+
+        """
+        super().__init__()
+
+        if not 0 <= skip_dd_threshold <= 1:
+            raise ValueError(f"skip_dd_threshold must be in [0, 1], but is {skip_dd_threshold}")
+
+        if min_duration is None:
+            if target.dt is None:
+                min_duration = 0
+            else:
+                min_duration = 2 * _gate_length_variance(target) / target.dt
+
+        self._min_duration = min_duration
+        self._skip_reset_qubits = skip_reset_qubits
+        self._skip_dd_threshold = skip_dd_threshold
+        self._target = target
+        self._coupling_map = target.build_coupling_map()  # build once and re-use for performance
+        self._pulse_alignment = (
+            target.pulse_alignment if pulse_alignment is None else pulse_alignment
+        )
+        self._coloring_strategy = coloring_strategy
+        self._sequence_generator = WalshHadamardSequence()
+
+        # Use PadDelay to insert Delay operations into the DAG before running this pass.
+        # This could be integrated into this pass as well, saving one iteration over the DAG,
+        # but this would repeat logic and is currently no bottleneck.
+        self.requires = [PadDelay(target=target)]
+
+    def run(self, dag: DAGCircuit) -> DAGCircuit:
+        # check the schedule analysis was run
+        if "node_start_time" not in self.property_set:
+            raise RuntimeError(
+                "node_start_time not found in the pass manager's property set. "
+                "Please run a scheduling pass before calling ContextAwareDynamicalDecoupling."
+            )
+
+        # 1) Find all delay instructions that qualify. We split them into explicit
+        # "begin" and "end" events and sort them.
+        sorted_delay_events = self._get_sorted_delays(dag)
+
+        # 2) Identify adjacent delays by overlap in time and adjacency on device.
+        qubit_map = {bit: idx for idx, bit in enumerate(dag.qubits)}  # map qubits to their indices
+        adjacent_delay_blocks = self._collect_adjacent_delay_blocks(sorted_delay_events, qubit_map)
+        logger.debug("adjacent delay blocks: %s", pformat(adjacent_delay_blocks))
+
+        # 3) Split the blocks into multi-qubit delay layers
+        merged_delays = []
+        for block in adjacent_delay_blocks:
+            merged_delays_in_block = self._split_adjacent_block(block, qubit_map)
+            merged_delays += merged_delays_in_block
+
+        # 4) For each multi-qubit delay, find a coloring (that takes into account neighboring
+        # CX or ECR gates) and get the DD sequence
+        all_delays = set()  # keep a list of all delays to iterate over them easily later
+
+        for merged_delay in merged_delays:
+            all_delays.update(merged_delay.ops)
+
+            # get coloring inside the n-qubit delay
+            coloring = self._get_wire_coloring(dag, merged_delay)
+
+            # get the DD sequence on the qubit
+            duration = merged_delay.end - merged_delay.start
+            checked_durations_cache = set()
+            for op in merged_delay.ops:
+                dd_circuit, start_times = self._get_dd_sequence(
+                    coloring[op.index],
+                    op.index,
+                    duration,
+                    checked_durations_cache,
+                )
+                op.replacement.compose(dd_circuit, inplace=True, copy=False)
+                if len(op.start_times) == 0:
+                    op.start_times += start_times
+                else:
+                    start_times = [op.start_times[-1] + time for time in start_times]
+
+                op.start_times += start_times
+
+        # 5) Replace each delay operation with its individual DD sequence
+        qubit_map = dict(enumerate(dag.qubits))
+        for delay in all_delays:
+            # replace it with its stored replacement
+            as_dag = circuit_to_dag(delay.replacement)
+            node_ids = [node._node_id for node in as_dag.topological_op_nodes()]
+            id_map = dag.substitute_node_with_dag(
+                delay.op, as_dag, {as_dag.qubits[0]: qubit_map[delay.index]}
+            )
+
+            # update node start times
+            for node_id, start_time in zip(node_ids, delay.start_times):
+                self.property_set["node_start_time"][id_map[node_id]] = start_time
+
+        return dag
+
+    def get_orthogonal_sequence(self, order: int) -> tuple[list[float], list[Gate]]:
+        """Return a DD sequence of given order, where different orders are orthogonal."""
+        spacing = self._sequence_generator.get_sequence(order)
+        return spacing, [XGate() for _ in spacing[:-1]]
+
+    def _get_sorted_delays(
+        self,
+        dag: DAGCircuit,
+    ) -> list[DelayEvent]:
+        """Get sorted DelayEvent objects for all eligible delay operations."""
+
+        def is_after_reset(node):
+            if not self._skip_reset_qubits:
+                return False  # if we do not skip reset qubits, we always want to use the delay
+
+            predecessor = next(dag.predecessors(node))  # a delay has one 1 predecessor
+            return isinstance(predecessor, DAGInNode) or isinstance(predecessor.op, Reset)
+
+        qubit_map = {bit: index for index, bit in enumerate(dag.qubits)}
+
+        eligible_delays = [
+            DelayEvent(
+                event_type,
+                (
+                    start_time
+                    if event_type == EventType.BEGIN
+                    else start_time + self._duration(node, qubit_map)
+                ),
+                node,
+            )
+            for node, start_time in self.property_set["node_start_time"].items()
+            if (
+                node.op.name == "delay"
+                and not is_after_reset(node)
+                and self._duration(node, qubit_map) > self._min_duration
+            )
+            for event_type in (EventType.BEGIN, EventType.END)
+        ]
+
+        sorted_events = sorted(eligible_delays, key=DelayEvent.sort_key)
+        logger.debug("Sorted delay events: %s", sorted_events)
+
+        return sorted_events
+
+    def _get_wire_coloring(self, dag: DAGCircuit, merged_delay: MultiDelayOp) -> dict[int, int]:
+        """Find a wire coloring for a multi-delay operation.
+
+        This function returns a dictionary that includes the coloring (as int) for the indices in the
+        ``merged_delay`` as ``{index: color}`` pairs. Spectator qubits are handled by assigning
+        neighboring qubits with a CX or ECR a color (0 if control, 1 if target) and including them
+        in the coloring problem.
+        """
+        # get neighboring wires, for which we will give initial colors
+        neighbors = set()
+        index_map = dict(enumerate(dag.qubits))
+        qubit_map = {bit: index for index, bit in enumerate(dag.qubits)}
+
+        for delay_op in merged_delay.ops:
+            # use coupling_map.graph.neighbors_undirected once Qiskit/rustworkx#1254 is in a release
+            new_neighbors = {
+                i
+                for i in range(dag.num_qubits())
+                if self._coupling_map.distance(i, delay_op.index) == 1
+            }
+            # do not add indices that are already in the merged delay
+            neighbors.update(new_neighbors.difference(merged_delay.indices))
+
+        # build a subgraph we will apply the coloring function on
+        wires = sorted(neighbors.union(merged_delay.indices))
+        subgraph = self._coupling_map.graph.subgraph(list(wires)).to_undirected()
+        glob2loc = dict(zip(wires, subgraph.node_indices()))
+        preset_coloring = {index: None for index in subgraph.node_indices()}
+
+        # find the neighbor wires and check if ctrl/tgt spectator
+        for wire in neighbors:
+            for op_node in dag.nodes_on_wire(dag.qubits[wire], only_ops=True):
+                # check if the operation occurs during the delay
+                op_start = self.property_set["node_start_time"][op_node]
+                op_end = op_start + self._duration(op_node, qubit_map)
+                if (
+                    isinstance(op_node.op, (CXGate, ECRGate))
+                    and op_start < merged_delay.end
+                    and op_end > merged_delay.start
+                ):
+                    # set coloring to 0 if ctrl, and to 1 if tgt
+                    ctrl, tgt = op_node.qargs
+                    if index_map[wire] == ctrl:
+                        preset_coloring[glob2loc[wire]] = 0
+                    if index_map[wire] == tgt:
+                        preset_coloring[glob2loc[wire]] = 1
+
+        local_coloring = rx.graph_greedy_color(
+            subgraph,
+            preset_color_fn=lambda loc: preset_coloring[loc],
+            strategy=self._coloring_strategy,
+        )
+
+        # map the local indices of the subgraph back to our DAG indices
+        loc2glob = dict(zip(subgraph.node_indices(), wires))
+        coloring = {loc2glob[loc]: color for loc, color in local_coloring.items()}
+
+        # for debugging purposes, print the coloring of each block
+        logger.debug("Coloring for block %s: %s", merged_delay, coloring)
+
+        return coloring
+
+    def _get_dd_sequence(
+        self,
+        order: int,
+        index: int,
+        duration: int,
+        checked_durations_cache: set[int],
+    ) -> tuple[QuantumCircuit, list[int]]:
+        """Get a DD sequence of specified order on qubit with a given index.
+
+        Takes the gate durations, the pulse alignment and a set to cache, which gate lengths
+        we've already checked to match the pulse alignment. Returns the DD sequence and the
+        node start times as list.
+        """
+        instruction_durations = self._target.durations()
+        # check the X gate on the active qubit is compatible with pulse alignment
+        if index not in checked_durations_cache:
+            x_duration = instruction_durations.get("x", index)
+            if x_duration % self._pulse_alignment != 0:
+                raise TranspilerError(
+                    f"X gate length on qubit with index {index} is {x_duration} which is not "
+                    f"an integer multiple of the pulse alignment {self._pulse_alignment}."
+                )
+            checked_durations_cache.add(index)
+
+        spacing, dd_sequence = self.get_orthogonal_sequence(order=order)
+
+        # check if DD can be applied or if there is not enough time
+        dd_sequence_duration = sum(instruction_durations.get(gate, index) for gate in dd_sequence)
+        slack = duration - dd_sequence_duration
+        slack_fraction = slack / duration
+        if 1 - slack_fraction >= self._skip_dd_threshold:  # dd doesn't fit
+            seq = QuantumCircuit(1)
+            seq.delay(duration, 0)
+            return seq, [0]
+
+        # compute actual spacings in between the delays, taking into account
+        # the pulse alignment restriction of the hardware
+        taus = self._constrain_spacing(spacing, slack)
+
+        # apply the DD gates
+        # tau has one more entry than the gate sequence
+        start_times = []
+        time = 0  # track the node start time
+        seq = QuantumCircuit(1)  # if the DD sequence has a global phase, add it here
+        for tau, gate in itertools.zip_longest(taus, dd_sequence):
+            if tau > 0:
+                seq.delay(tau, 0)
+                start_times.append(time)
+                time += tau
+            if gate is not None:
+                seq.append(gate, [0])
+                start_times.append(time)
+                time += instruction_durations.get(gate, index)
+
+        return seq, start_times
+
+    def _constrain_spacing(self, spacing, slack):
+        def _constrained_length(values):
+            return self._pulse_alignment * np.floor(values / self._pulse_alignment)
+
+        taus = _constrained_length(slack * np.asarray(spacing))
+        unused_slack = slack - sum(taus)  # unused, due to rounding to int multiples of dt
+        middle_index = int((len(taus) - 1) / 2)  # arbitrary: redistribute to middle
+        to_middle = _constrained_length(unused_slack)
+        taus[middle_index] += to_middle  # now we add up to original delay duration
+        if unused_slack - to_middle:
+            taus[-1] += unused_slack - to_middle
+
+        return taus
+
+    def _collect_adjacent_delay_blocks(
+        self,
+        sorted_delay_events: list[DelayEvent],
+        qubit_map: dict[Qubit, int],
+    ) -> list[AdjacentDelayBlock]:
+        """Collect delay events into adjacent blocks.
+
+        Events in an adjacent block are overlapping in time and adjacent on the device.
+        See also the dataclass ``AdjacentDelayBlock`` for more information.
+
+        Args:
+            sorted_delay_events: All eligible delay operations, sorted by time and type.
+            qubit_map: A map from qubit instance to qubit index.
+
+        Returns:
+            A list of adjacent delay blocks.
+        """
+        open_delay_blocks = []
+        closed_delay_blocks = []
+
+        def _open_delay_block(delay_event):
+            open_delay_blocks.append(
+                AdjacentDelayBlock(
+                    events=[delay_event], active_qubits=set(delay_event.op_node.qargs)
+                )
+            )
+            return open_delay_blocks[-1]
+
+        def _update_delay_block(open_delay_block, delay_event):
+            """Add another delay event to an existing block to either extend or close it."""
+            open_delay_block.events.append(delay_event)
+
+            # at this point we know that delay_event.op_node.qargs is active
+            open_delay_block.active_qubits -= set(delay_event.op_node.qargs)
+
+            if not open_delay_block.active_qubits:
+                open_delay_blocks.remove(open_delay_block)
+                closed_delay_blocks.append(open_delay_block)
+
+        def _combine_delay_blocks(delay_blocks):
+            survivor, *doomed = delay_blocks
+
+            for doomed_delay_group in doomed:
+                # Add events and qubits from doomed block to survivor.
+                if logger.isEnabledFor(logging.DEBUG):
+                    if survivor.active_qubits.intersection(doomed_delay_group.active_qubits):
+                        logger.debug("More than one open delay on a qubit?")
+
+                survivor.events.extend(doomed_delay_group.events)
+                survivor.active_qubits.update(doomed_delay_group.active_qubits)
+
+                open_delay_blocks.remove(doomed_delay_group)
+                survivor.events.sort(key=DelayEvent.sort_key)  # Maintain sorted event order
+
+        for delay_event in sorted_delay_events:
+            # This could be avoided by keeping a map of device qubit to open
+            # block and only considering neighbors of current event.
+            # use coupling_map.graph.neighbors_undirected once Qiskit/rustworkx#1254 is in a release
+            adjacent_open_delay_blocks = [
+                open_delay
+                for open_delay in open_delay_blocks
+                if any(
+                    self._coupling_map.distance(
+                        qubit_map[delay_event.op_node.qargs[0]], qubit_map[open_delay_qubit]
+                    )
+                    <= 1
+                    for open_delay_qubit in open_delay.active_qubits
+                )
+            ]
+
+            if delay_event.type == EventType.BEGIN:
+                # If crossing a begin edge, check if there are any open delays that are adjacent.
+                # If so, add current event to that group.
+
+                if len(adjacent_open_delay_blocks) == 0:
+                    # Make a new delay block
+                    _open_delay_block(delay_event)
+                else:
+                    # Make a new block and combine that with adjacent open blocks
+                    new_block = _open_delay_block(delay_event)
+                    _combine_delay_blocks(adjacent_open_delay_blocks + [new_block])
+
+            else:
+                if logger.isEnabledFor(logging.DEBUG):
+                    # If crossing a end edge, remove this qubit from the actively delaying qubits"
+                    if len(adjacent_open_delay_blocks) != 1:
+                        logger.debug("Closing edge w/o an open delay?")
+
+                _update_delay_block(adjacent_open_delay_blocks[0], delay_event)
+
+        # log the open delays and the number of closed delays
+        logger.debug("-- post collect")
+        logger.debug("open delays: %s", open_delay_blocks)
+        logger.debug("len(closed delays): %s", len(closed_delay_blocks))
+
+        # validate the results, there should be no open delays and all active qubits
+        # should be accounted for
+        if logger.isEnabledFor(logging.DEBUG):
+            if len(open_delay_blocks) > 0:
+                logger.debug("Failed to close all open delays.")
+
+            for closed_delay in closed_delay_blocks:
+                if len(closed_delay.active_qubits) > 0:
+                    logger.debug("Failed to remove active qubits on closed delay %s.", closed_delay)
+
+                closed_delay.validate()
+
+        return closed_delay_blocks
+
+    def _split_adjacent_block(
+        self,
+        adjacent_block: AdjacentDelayBlock,
+        qubit_map: dict[Qubit, int],
+    ) -> list[MultiDelayOp]:
+        """Split adjacent delay blocks in concurrent layers of maximum width.
+
+        This code performs the following steps:
+
+            1) Find the atomic time windows where delays can begin and end.
+            2) Find which connected qubit components are jointly idle during the windows.
+            3) Merge time-adjacent qubit components.
+
+        """
+        # 1) Find times at which a new delay event is happening and we might have to break up
+        # the adjacent delays
+        breakpoints = []
+        # a dictionary of {qubit_index: [delay_op1, delay_op2, ...]}
+        all_delays = defaultdict(list)
+
+        for event in adjacent_block.events:
+            index = qubit_map[event.op_node.qargs[0]]  # delays are 1q ops
+            if event.type == EventType.BEGIN:
+                delay = DelayOp(start=event.time, index=index, op=event.op_node)
+                all_delays[index].append(delay)
+            else:  # find which delay to close
+                all_delays[index][-1].end = event.time
+
+            if len(breakpoints) == 0 or event.time > breakpoints[-1]:
+                breakpoints.append(event.time)
+
+        # 2) Find which delays are active during which time windows, where a time window
+        # is defined as (breakpoints[i], breakpoints[i + 1])
+        active_delays = {}  # {window: [group1, group2]} where each group = (index1, index2, ..)
+        op_map = {}  # {(qubit_index, window): delay operation as DAGOpNode}
+        windows = list(zip(breakpoints[:-1], breakpoints[1:]))
+        for window in windows:
+            active = []
+
+            # check which delays are active during the window
+            # this could be e.g. [0, 1, 2, 5, 6, 9]
+            for index, delays in all_delays.items():
+                for delay in delays:
+                    # since the windows are atomic, we have three cases:
+                    # (1) the delay starts at the time window, (2) it ends with it, (3) it contains it
+                    if delay.start <= window[0] and delay.end >= window[1]:
+                        active.append(index)
+                        delay.add_window(window)
+                        op_map[(index, window)] = delay
+
+            # check which are adjacent
+            # on a linear topology, we would get [[0, 1, 2], [5, 6], [9]]
+            visited = defaultdict(lambda: False)
+            grouped = []
+            for start_index in active:
+                if visited[start_index]:
+                    continue
+
+                active_neighbors = {start_index}
+                _dfs(start_index, self._coupling_map, active_neighbors, active)
+
+                for index in active_neighbors:
+                    visited[index] = True
+
+                group = tuple(sorted(active_neighbors))  # must be sorted to merge later
+                grouped.append(group)
+
+            # sanity check: groups must be disjoint
+            if logger.isEnabledFor(logging.DEBUG):
+                for i, g1 in enumerate(grouped):
+                    for g2 in grouped[i + 1 :]:
+                        if len(set(g1).intersection(g2)) > 0:
+                            logger.debug("Groups not disjoint: %s and %s.", g1, g2)
+
+            active_delays[window] = grouped
+
+        # 3) Merge time-adjacent active delays
+        merged_delays = []
+        open_groups = {}
+        for window in windows:
+            next_groups = active_delays[window]
+
+            # check which opened groups are still active in this window
+            next_open_groups = {}
+            for open_group, delay in open_groups.items():
+                if open_group in next_groups:
+                    # extend the delay operation and remove a breakpoint
+                    delay.end = window[1]
+                    for op in delay.ops:
+                        op.breakpoints.remove(window[0])
+
+                    if logger.isEnabledFor(logging.DEBUG):
+                        involved_ops = {op_map[(index, window)] for index in open_group}
+                        if len(involved_ops.difference(delay.ops)) > 0:
+                            logger.debug("Not all involved operations are part of the joint delay.")
+
+                    next_open_groups[open_group] = delay
+                else:
+                    # the group has not been extended, close it!
+                    merged_delays.append(delay)
+
+            # add the new open groups
+            for group in next_groups:
+                if group not in open_groups:
+                    involved_ops = {op_map[(index, window)] for index in group}
+                    next_open_groups[group] = MultiDelayOp(
+                        start=window[0], end=window[1], indices=group, ops=involved_ops
+                    )
+
+            open_groups = next_open_groups
+
+        # add all from the final layer
+        for open_group, delay in open_groups.items():
+            merged_delays.append(delay)
+
+        logger.debug("Split adjacent delay block into %s", "\n".join(map(str, merged_delays)))
+
+        return merged_delays
+
+    def _duration(self, node: DAGOpNode, qubit_map: dict[Qubit, int]) -> float:
+        # this is cached on the target, so we can repeatedly call it w/o penalty
+        instruction_durations = self._target.durations()
+        indices = [qubit_map[bit] for bit in node.qargs]
+
+        return instruction_durations.get(node.op, indices)
+
+
+class EventType(Enum):
+    """Delay event type, which is either begin or end."""
+
+    BEGIN = 0
+    END = 1
+
+
+@dataclass
+class DelayEvent:
+    """Represent a single-qubit delay event, which is either begin or end of a delay instruction."""
+
+    type: EventType
+    time: int  # Time in this circuit, in dt
+    op_node: DAGOpNode  # The node for the circuit delay
+
+    @staticmethod
+    def sort_key(event: DelayEvent) -> tuple(int, int):
+        """Sort events, first by time then by type ('end' events come before 'begin' events)."""
+        return (
+            event.time,  # Sort by event time
+            0 if event.type == EventType.END else 1,  # With 'end' events before 'begin'
+        )
+
+
+@dataclass
+class DelayOp:
+    """Represent a delay operation."""
+
+    start: int
+    index: int
+    op: DAGOpNode  # the circuit op node this represents
+    end: int | None = None  # None means currently unknown
+    breakpoints: list[int] = field(
+        default_factory=list
+    )  # timepoints at which the delay op is split
+    # circuit with which we will replace this delay
+    start_times: list[int] = field(default_factory=list)
+    replacement: QuantumCircuit = field(default_factory=lambda: QuantumCircuit(1))
+
+    def __hash__(self):
+        return hash(self.op)
+
+    def add_window(self, window: tuple[int, int]):
+        """Add a time window to the delay op.
+
+        This means the delay is active during this window and we add a potential breakpoint.
+        """
+        if self.end is None:
+            raise ValueError("Cannot add a window if DelayOp.end is None. Please set it.")
+
+        start, end = window
+        if self.start < start and start not in self.breakpoints:
+            self.breakpoints.append(start)
+        if self.end > end and end not in self.breakpoints:
+            self.breakpoints.append(end)
+
+        self.breakpoints = sorted(self.breakpoints)
+
+
+@dataclass
+class MultiDelayOp:
+    """A multi-qubit delay operation."""
+
+    start: int
+    end: int
+    indices: list[int]
+    ops: set[DelayOp]
+    replacing: set[DAGOpNode] = field(default_factory=set)
+
+    def __str__(self) -> str:
+        return f"MultiDelay({self.start}:{self.end} on {self.indices})"
+
+
+@dataclass
+class AdjacentDelayBlock:
+    """Group of circuit delays which are collectively adjacent in time and on device.
+
+    For example, here the 3 delay operations on q0, q1 and q2 form an adjacent delay block.
+
+        q0: -██████---------  |  qubits q0,q1,q2 have adjacent delay
+        q1: ------███████---  |  operations, since the delay operations
+        q2: --█████████-----  |  all overlap
+        q3: -----------████-  -> this delay starts when delay on q2 ends, so they have no overlap
+        q4: ----████--------  -> this clearly has no overlap with something else
+
+    """
+
+    events: list[DelayEvent]
+    active_qubits: set[Qubit]
+
+    def validate(self, log: bool = True) -> None:
+        """Validate the list of delay events in the adjacent block.
+
+        Args:
+            log: If ``True`` log invalid blocks on DEBUG level. Otherwise raise an error if the
+                block is invalid.
+
+        Raises:
+            RuntimeError: If the blocks are not ordered by time and event type.
+        """
+
+        def notify(msg, *args):
+            if log:
+                logger.debug(msg, *args)
+            else:
+                raise RuntimeError(msg.format(*args))
+
+        for idx, event in enumerate(self.events[:-1]):
+            if event.time > self.events[idx + 1].time:
+                notify("adjacent_delay_block.events not ordered by time")
+
+            if event.time == self.events[idx + 1].time:
+                # At same time, can either be ('begin', 'begin'), ('end', 'begin') or ('end', 'end')
+                if (event.type, self.events[idx + 1].type) == (EventType.BEGIN, EventType.END):
+                    notify(
+                        "Events in the AdjacentDelayBlock are not correctly sorted by "
+                        "event type. At same time, we can have either of (begin, begin), "
+                        "(end, begin) or (end, end). This happened at time %s.",
+                        event.time,
+                    )
+
+
+def _dfs(qubit, cmap: CouplingMap, visited, active_qubits):
+    """Depth-first search to get the widest group of idle qubits during a given time frame."""
+    # use coupling_map.graph.neighbors_undirected once Qiskit/rustworkx#1254 is in a release
+    neighbors = {other for other in active_qubits if cmap.distance(qubit, other) == 1}
+    for neighbor in neighbors:
+        if neighbor in active_qubits and neighbor not in visited:
+            visited.add(neighbor)
+            _dfs(neighbor, cmap, visited, active_qubits)
+
+
+class WalshHadamardSequence:
+    """Get Walsh-Hadamard sequences for DD up to arbitrary order."""
+
+    def __init__(self, max_order: int = 5):
+        """
+        Args:
+            max_order: The maximal order for which the sequences are computed.
+        """
+        # these are set in set_max_order
+        self.sequences = None
+        self.max_order = None
+
+        self.set_max_order(max_order)
+
+    def set_max_order(self, max_order: int) -> None:
+        """Set the maximal available order."""
+        if self.max_order is not None:
+            if max_order <= self.max_order:
+                return
+
+        # get the dimension of the transformation matrix we need,
+        # this is given by the smallest power of 2 that includes max_order
+        num_krons = int(np.ceil(np.log2(max_order + 1)))
+        self.max_order = 2**num_krons - 1
+
+        rows = _get_transformation_matrix(num_krons).tolist()
+        distances = [_bitflips_to_timings(row) for row in rows]
+        num_flips = [len(distance) - 1 for distance in distances]
+
+        # sort by the number of flips and throw out the first one,
+        # which corresponds to the 000... bit-sequence, i.e., no flip
+        indices = np.argsort(num_flips)[1:]
+
+        self.sequences = [distances[i] for i in indices]
+
+    def get_sequence(self, order: int) -> list[float]:
+        """Get the Walsh-Hadamard sequence of given order (starts at 0)."""
+        if order > self.max_order:
+            self.set_max_order(order)
+
+        return self.sequences[order]
+
+
+def _get_transformation_matrix(n):
+    """Get a 2^n x 2^n Walsh-Hadamard matrix with elements in [0, 1]."""
+
+    from qiskit.circuit.library import HGate
+
+    had = np.array(HGate()).real * np.sqrt(2)
+
+    def recurse(matrix, m):
+        # we build the matrix recursively, adding one Hadamard kronecker product per recursion
+        if m == 1:
+            # finally, map to [0, 1] by the mapping (1 - H) / 2
+            return ((1 - matrix) / 2).astype(int)
+
+        return recurse(np.kron(had, matrix), m - 1)
+
+    return recurse(had, n)
+
+
+def _bitflips_to_timings(row):
+    num = len(row)
+    distances = []
+    count = 0
+    last = 0  # start in no flip state
+    for el in row:
+        if el == last:
+            count += 1
+        else:
+            distances.append(count / num)
+            last = el
+            count = 1
+
+    distances.append(count / num)
+
+    if len(distances) % 2 == 0:
+        return distances + [0]
+
+    return distances
+
+
+def _gate_length_variance(target: Target) -> float:
+    max_length, min_length = None, None
+
+    for gate, properties in target.items():
+        if gate not in ["cx", "cz", "ecr"]:
+            continue
+
+        for prop in properties.values():
+            duration = prop.duration
+            if max_length is None or max_length < duration:
+                max_length = duration
+            if min_length is None or min_length > duration:
+                min_length = duration
+
+    # it could be that there are no 2q gates available, in this
+    # case we just return 0, which will mean we join all idle times
+    if max_length is None or min_length is None:
+        return 0
+
+    return max_length - min_length