Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Speedup constant factors in LookaheadSwap #8068

Merged
merged 4 commits into from
Jun 21, 2022
Merged
Show file tree
Hide file tree
Changes from 3 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
218 changes: 122 additions & 96 deletions qiskit/transpiler/passes/routing/lookahead_swap.py
Original file line number Diff line number Diff line change
Expand Up @@ -12,10 +12,11 @@

"""Map input circuit onto a backend topology via insertion of SWAPs."""

import collections
import copy
import logging
from copy import deepcopy
import math

from qiskit.circuit.quantumregister import QuantumRegister
from qiskit.circuit.library.standard_gates import SwapGate
from qiskit.transpiler.basepasses import TransformationPass
from qiskit.transpiler.exceptions import TranspilerError
Expand All @@ -25,6 +26,25 @@
logger = logging.getLogger(__name__)


_Step = collections.namedtuple("_Step", ("state", "swaps_added", "gates_mapped", "gates_remaining"))
"""Describes one possible step in the lookahead process.

The fields are:

state (_SystemState): The current state of the system, including its virtual-to-physical layout.
swaps_added (list): List of qargs of swap gates introduced.
gates_mapped (list): Gates that were mapped, including added SWAPs.
gates_remaining (list): Gates that could not be mapped.
"""

_SystemState = collections.namedtuple(
"_SystemState",
("layout", "coupling_map", "register", "swaps"),
# The None default applies to the right-most element, i.e. `swaps`.
defaults=(None,),
)


class LookaheadSwap(TransformationPass):
"""Map input circuit onto a backend topology via insertion of SWAPs.

Expand Down Expand Up @@ -100,21 +120,20 @@ def run(self, dag):
f"available device qubits ({number_of_available_qubits})."
)

canonical_register = dag.qregs["q"]
trivial_layout = Layout.generate_trivial_layout(canonical_register)
current_layout = trivial_layout.copy()
register = dag.qregs["q"]
current_state = _SystemState(
Layout.generate_trivial_layout(register), self.coupling_map, register
)

mapped_gates = []
ordered_virtual_gates = list(dag.serial_layers())
gates_remaining = ordered_virtual_gates.copy()
gates_remaining = list(dag.serial_layers())

while gates_remaining:
logger.debug("Top-level routing step: %d gates remaining.", len(gates_remaining))

best_step = _search_forward_n_swaps(
current_layout,
current_state,
gates_remaining,
self.coupling_map,
self.search_depth,
self.search_width,
)
Expand All @@ -126,108 +145,90 @@ def run(self, dag):

logger.debug(
"Found best step: mapped %d gates. Added swaps: %s.",
len(best_step["gates_mapped"]),
best_step["swaps_added"],
len(best_step.gates_mapped),
best_step.swaps_added,
)

current_layout = best_step["layout"]
gates_mapped = best_step["gates_mapped"]
gates_remaining = best_step["gates_remaining"]
current_state = best_step.state
gates_mapped = best_step.gates_mapped
gates_remaining = best_step.gates_remaining

mapped_gates.extend(gates_mapped)

if self.fake_run:
self.property_set["final_layout"] = current_layout
self.property_set["final_layout"] = current_state.layout
return dag

# Preserve input DAG's name, regs, wire_map, etc. but replace the graph.
mapped_dag = dag.copy_empty_like()

for node in mapped_gates:
mapped_dag.apply_operation_back(op=node.op, qargs=node.qargs, cargs=node.cargs)

return mapped_dag


def _search_forward_n_swaps(layout, gates, coupling_map, depth, width):
def _search_forward_n_swaps(state, gates, depth, width):
"""Search for SWAPs which allow for application of largest number of gates.

Args:
layout (Layout): Map from virtual qubit index to physical qubit index.
state (_SystemState): The ``namedtuple`` collection containing the state of the physical
system. This includes the current layout, the coupling map, the canonical register and
the possible swaps available.
gates (list): Gates to be mapped.
coupling_map (CouplingMap): CouplingMap of the target backend.
depth (int): Number of SWAP layers to search before choosing a result.
width (int): Number of SWAPs to consider at each layer.
Returns:
optional(dict): Describes solution step found. If None, no swaps leading
to an improvement were found. Keys:
layout (Layout): Virtual to physical qubit map after SWAPs.
swaps_added (list): List of qargs of swap gates introduced.
gates_remaining (list): Gates that could not be mapped.
gates_mapped (list): Gates that were mapped, including added SWAPs.

Optional(_Step): Describes the solution step found. If ``None``, no swaps leading to an
improvement were found.
"""
gates_mapped, gates_remaining = _map_free_gates(layout, gates, coupling_map)

base_step = {
"layout": layout,
"swaps_added": [],
"gates_mapped": gates_mapped,
"gates_remaining": gates_remaining,
}
if state.swaps is None:
# Include symmetric couplings (e.g [0,1] and [1,0]) as one swap.
state = state._replace(
swaps={((a, b) if a < b else (b, a)) for a, b in state.coupling_map.get_edges()}
)
gates_mapped, gates_remaining = _map_free_gates(state, gates)
base_step = _Step(state, [], gates_mapped, gates_remaining)

if not gates_remaining or depth == 0:
return base_step

# Include symmetric 2q gates (e.g coupling maps with both [0,1] and [1,0])
# as one available swap.
possible_swaps = {tuple(sorted(edge)) for edge in coupling_map.get_edges()}

def _score_swap(swap):
"""Calculate the relative score for a given SWAP."""
trial_layout = layout.copy()
trial_layout.swap(*swap)
return _calc_layout_distance(gates, coupling_map, trial_layout)

ranked_swaps = sorted(possible_swaps, key=_score_swap)
ranked_swaps = sorted(
(_score_state_with_swap(swap, state, gates) for swap in state.swaps),
key=lambda x: x[0],
)
logger.debug(
"At depth %d, ranked candidate swaps: %s...",
depth,
[(swap, _score_swap(swap)) for swap in ranked_swaps[: width * 2]],
[(swap, score) for score, swap, _ in ranked_swaps[: width * 2]],
)

best_swap, best_step = None, None
for rank, swap in enumerate(ranked_swaps):
trial_layout = layout.copy()
trial_layout.swap(*swap)
next_step = _search_forward_n_swaps(
trial_layout, gates_remaining, coupling_map, depth - 1, width
)
best_swap, best_step, best_score = None, None, -math.inf
for rank, (_, swap, new_state) in enumerate(ranked_swaps):
next_step = _search_forward_n_swaps(new_state, gates_remaining, depth - 1, width)

if next_step is None:
continue

next_score = _score_step(next_step)
# ranked_swaps already sorted by distance, so distance is the tie-breaker.
if best_swap is None or _score_step(next_step) > _score_step(best_step):
if next_score > best_score:
logger.debug(
"At depth %d, updating best step: %s (score: %f).",
depth,
[swap] + next_step["swaps_added"],
_score_step(next_step),
[swap] + next_step.swaps_added,
next_score,
)
best_swap, best_step = swap, next_step
best_swap, best_step, best_score = swap, next_step, next_score

if (
rank >= min(width, len(ranked_swaps) - 1)
and best_step is not None
and (
len(best_step["gates_mapped"]) > depth
or len(best_step["gates_remaining"]) < len(gates_remaining)
len(best_step.gates_mapped) > depth
or len(best_step.gates_remaining) < len(gates_remaining)
or (
_calc_layout_distance(
best_step["gates_remaining"], coupling_map, best_step["layout"]
)
< _calc_layout_distance(gates_remaining, coupling_map, layout)
_calc_layout_distance(best_step.gates_remaining, best_step.state)
< _calc_layout_distance(gates_remaining, new_state)
)
)
):
Expand All @@ -239,24 +240,24 @@ def _score_swap(swap):
else:
return None

logger.debug("At depth %d, best_swap set: %s.", depth, [best_swap] + best_step["swaps_added"])

best_swap_gate = _swap_ops_from_edge(best_swap, layout)
return {
"layout": best_step["layout"],
"swaps_added": [best_swap] + best_step["swaps_added"],
"gates_remaining": best_step["gates_remaining"],
"gates_mapped": gates_mapped + best_swap_gate + best_step["gates_mapped"],
}
best_swap_gate = _swap_ops_from_edge(best_swap, state)
out = _Step(
best_step.state,
[best_swap] + best_step.swaps_added,
gates_mapped + best_swap_gate + best_step.gates_mapped,
best_step.gates_remaining,
)
logger.debug("At depth %d, best_swap set: %s.", depth, out.swaps_added)
return out


def _map_free_gates(layout, gates, coupling_map):
def _map_free_gates(state, gates):
"""Map all gates that can be executed with the current layout.

Args:
layout (Layout): Map from virtual qubit index to physical qubit index.
state (_SystemState): The physical characteristics of the system, including its current
layout and the coupling map.
gates (list): Gates to be mapped.
coupling_map (CouplingMap): CouplingMap for target device topology.

Returns:
tuple:
Expand All @@ -267,12 +268,13 @@ def _map_free_gates(layout, gates, coupling_map):

mapped_gates = []
remaining_gates = []
layout_map = state.layout._v2p

for gate in gates:
# Gates without a partition (barrier, snapshot, save, load, noise) may
# still have associated qubits. Look for them in the qargs.
if not gate["partition"]:
qubits = [n for n in gate["graph"].nodes() if isinstance(n, DAGOpNode)][0].qargs
qubits = _first_op_node(gate["graph"]).qargs

if not qubits:
continue
Expand All @@ -281,7 +283,7 @@ def _map_free_gates(layout, gates, coupling_map):
blocked_qubits.update(qubits)
remaining_gates.append(gate)
else:
mapped_gate = _transform_gate_for_layout(gate, layout)
mapped_gate = _transform_gate_for_system(gate, state)
mapped_gates.append(mapped_gate)
continue

Expand All @@ -291,10 +293,10 @@ def _map_free_gates(layout, gates, coupling_map):
blocked_qubits.update(qubits)
remaining_gates.append(gate)
elif len(qubits) == 1:
mapped_gate = _transform_gate_for_layout(gate, layout)
mapped_gate = _transform_gate_for_system(gate, state)
mapped_gates.append(mapped_gate)
elif coupling_map.distance(*(layout[q] for q in qubits)) == 1:
mapped_gate = _transform_gate_for_layout(gate, layout)
elif state.coupling_map.distance(layout_map[qubits[0]], layout_map[qubits[1]]) == 1:
mapped_gate = _transform_gate_for_system(gate, state)
mapped_gates.append(mapped_gate)
else:
blocked_qubits.update(qubits)
Expand All @@ -303,43 +305,67 @@ def _map_free_gates(layout, gates, coupling_map):
return mapped_gates, remaining_gates


def _calc_layout_distance(gates, coupling_map, layout, max_gates=None):
def _calc_layout_distance(gates, state, max_gates=None):
"""Return the sum of the distances of two-qubit pairs in each CNOT in gates
according to the layout and the coupling.
"""
if max_gates is None:
max_gates = 50 + 10 * len(coupling_map.physical_qubits)
max_gates = 50 + 10 * len(state.coupling_map.physical_qubits)

layout_map = state.layout._v2p
out = 0
for gate in gates[:max_gates]:
if not gate["partition"]:
continue
qubits = gate["partition"][0]
if len(qubits) == 2:
out += state.coupling_map.distance(layout_map[qubits[0]], layout_map[qubits[1]])
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not really something we should change here, but I typically avoid the distance() method and just access the inner distance matrix directly because besides the function call overhead you also have bounds checking which for something like this you know you're layout isn't going to return bounds outside the matrix. You could probably also use numpy's sum to speed this up (although when I've tried that in the past it wasn't really a noticeable difference).

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looking back, I think I was just doing a direct port of the code here - it previously used coupling_map.distance, just with an iterable unpacking of a generator expression. I unrolled the generator, but left the function call.

return out

return sum(
coupling_map.distance(*(layout[q] for q in gate["partition"][0]))
for gate in gates[:max_gates]
if gate["partition"] and len(gate["partition"][0]) == 2
)

def _score_state_with_swap(swap, state, gates):
"""Calculate the relative score for a given SWAP.

Returns:
float: the score of the given swap.
Tuple[int, int]: the input swap that should be performed.
_SystemState: an updated system state with the new layout contained.
"""
trial_layout = state.layout.copy()
trial_layout.swap(*swap)
new_state = state._replace(layout=trial_layout)
return _calc_layout_distance(gates, new_state), swap, new_state


def _score_step(step):
"""Count the mapped two-qubit gates, less the number of added SWAPs."""
# Each added swap will add 3 ops to gates_mapped, so subtract 3.
return len([g for g in step["gates_mapped"] if len(g.qargs) == 2]) - 3 * len(
step["swaps_added"]
)
return len([g for g in step.gates_mapped if len(g.qargs) == 2]) - 3 * len(step.swaps_added)


def _transform_gate_for_layout(gate, layout):
def _transform_gate_for_system(gate, state):
"""Return op implementing a virtual gate on given layout."""
mapped_op_node = deepcopy([n for n in gate["graph"].nodes() if isinstance(n, DAGOpNode)][0])
mapped_op_node = copy.copy(_first_op_node(gate["graph"]))

device_qreg = QuantumRegister(len(layout.get_physical_bits()), "q")
mapped_qargs = [device_qreg[layout[a]] for a in mapped_op_node.qargs]
device_qreg = state.register
layout_map = state.layout._v2p
mapped_qargs = [device_qreg[layout_map[a]] for a in mapped_op_node.qargs]
mapped_op_node.qargs = mapped_qargs

return mapped_op_node


def _swap_ops_from_edge(edge, layout):
def _swap_ops_from_edge(edge, state):
"""Generate list of ops to implement a SWAP gate along a coupling edge."""
device_qreg = QuantumRegister(len(layout.get_physical_bits()), "q")
device_qreg = state.register
qreg_edge = [device_qreg[i] for i in edge]

# TODO shouldn't be making other nodes not by the DAG!!
return [DAGOpNode(op=SwapGate(), qargs=qreg_edge, cargs=[])]


def _first_op_node(dag):
"""Get the first op node from a DAG."""
# This doesn't use `DAGCircuit.op_nodes` because that function always consumes the entire
# iterator to create a list, whereas we only need the first element.
return next(node for node in dag.nodes() if isinstance(node, DAGOpNode))
Comment on lines +367 to +371
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reading through the code I don't think this matters but this is the first op node based on insertion order not necessarily topological order was that your intent or did you want the first node from a topological ordering of the dag?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is just a direct port of the existing code - it previously did this exact thing but constructed the complete list just to take the first element. I was mostly trying to ensure the iteration order wasn't affect while just trying to speed it up.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok, that's what it looked like to me too but I wanted to confirm. The code just read a little weird to me because the "first" node could actually be something in the middle of the dag or at the end as it's dependent on insertion order.

10 changes: 10 additions & 0 deletions releasenotes/notes/speedup-lookahead-swap-4dd162fee2d25d10.yaml
Original file line number Diff line number Diff line change
@@ -0,0 +1,10 @@
---
features:
- |
The transpiler pass :class:`.LookaheadSwap` (used by :func:`.transpile` when
``routing_method="lookahead"``) has seen some performance improvements and
will now be approximately three times as fast. This is purely being more
efficient in its calculations, and does not change the complexity of the
algorithm. In most cases, a more modern routing algorithm like
:class:`.SabreSwap` (``routing_method="sabre"``) will be vastly more
performant.