Calculate relative, not absolute, scores in SabreSwap (#9012)

* Calculate relative, not absolute, scores in SabreSwap

This is a significant performance improvement for very wide (100q+)
circuits.

The previous `SabreSwap` algorithm would score each candidate swap by
applying the swap to the current layout, iterating through every element
in the front layer and extended sets summing the total distances of the
2q gates, and then undoing the swap.  However, in the front layer, a
given swap can affect at most two 2q operations.  This new form instead
scores each swap by calculating its total distance relative to if the
swap had not been made.  This means that the basic score is formed from
only one or two gates, no matter how many are in the front layer.

This is an algorithmic complexity improvement in the scoring for
volumetric circuits with respect to the number of qubits.  Such a
circuit with `n` qubits has `n / 2` gates in its front layer at all
times, and so (assuming a coupling map that expands by a constant
connectivity factor per qubit, like heavy hex) `k * n` swaps to score.
This means that choosing the best swap has quadratic complexity with the
original Sabre scoring algorithm.  With this new algorithm, the score
for a given swap is calculated in constant time, so choosing the best is
instead linear.  In practice, I did not see all these improvements at
the scales I tested at, but I did see significant improvements - a 1081q
heavy-hex quantum-volume circuit at depth 5 was swap-mapped on my
machine in 25s with this commit compared to 100s before it.

The principal change is the structs `FrontLayer` and `ExtendedSet`,
which combine constant-time hash-set insertions, lookups and removals
with vectors to enable constant-time lookup of the affected qubits.
`FrontLayer` now only ever holds currently unroutable 2q operations;
routable operations are immediately placed into the output structures
as soon as a new 2q gate is routed.  This routing is also done by
exploiting that a swap can only affect two previously unroutable gates
in the front layer; we just walk forwards along the outbound edges from
those two nodes, adding any operations that become routable.  This
avoids another linear scan through the whole front layer after each
swap, although in practice this has less of a speed-up effect, because
it already wasn't quadratic.

In theory, this change does not affect how any swap is scored relative
to any other with the same front layer and extended set (though the
scores used in the comparison do change).  In order to precisely match
the current implementation (and ensure reproducibility from a given
seed), this tracks the insertion order of nodes into the front layer,
including after removals.

This commit completely modifies the internals of the Rust components of
Sabre, although the actual algorithm is largely unchanged, aside from
the scoring difference.  Various parts of the implementation do change
for efficiency, though.

This commit maintains RNG compatibility with the previous Rust
implementation in most cases.  It is possible in some circuits for
floating-point differences to cause different output, when several swaps
are at the minimum score, but plus/minus 1ULP.  This happens in both the
old and new forms of the implementation, but _which_ of the minimal
swaps get the minus-1ULP score varies between them, and consequently
affects the swap choice.  In fairly extensive testing, this appears to
be the only mechanism for differences; I've verified that the
release-valve mechanism and predecessor-requirement tracking function
identically to before.  The resultant scores - relative for "basic" and
"lookahead", absolute for "decay" - are in practice within 2ULP of the
old algorithm's.

In maintaining RNG compatibility, this commit leaves several further
speed-ups on the table.  There is additional memory usage and tracking
to maintain some required iteration orders, and some reordering checks
that are not strictly necessary any more.  Further, the sorting stages
at the ends of the swap-choosing functions (to maintain repeatability)
can be made redundant now, since some hash-set iteration (which is
effectively an uncontrolled randomisation per run) is no longer
required.  These will be addressed in follow-ups.

* Fix lint

Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

qiskit/transpiler/passes/routing/sabre_swap.py

@@ -223,8 +223,7 @@ def run(self, dag):
                     cargs,
                 )
             )
-        front_layer = np.asarray([x._node_id for x in dag.front_layer()], dtype=np.uintp)
-        sabre_dag = SabreDAG(len(dag.qubits), len(dag.clbits), dag_list, front_layer)
+        sabre_dag = SabreDAG(len(dag.qubits), len(dag.clbits), dag_list)
         swap_map, gate_order = build_swap_map(
             len(dag.qubits),
             sabre_dag,

src/sabre_swap/layer.rs

@@ -0,0 +1,295 @@
+// This code is part of Qiskit.
+//
+// (C) Copyright IBM 2022
+//
+// 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.
+
+use hashbrown::HashMap;
+use ndarray::prelude::*;
+use retworkx_core::petgraph::prelude::*;
+
+use crate::nlayout::NLayout;
+
+/// A container for the current non-routable parts of the front layer.  This only ever holds
+/// two-qubit gates; the only reason a 0q- or 1q operation can be unroutable is because it has an
+/// unsatisfied 2q predecessor, which disqualifies it from being in the front layer.
+pub struct FrontLayer {
+    /// Map of the (index to the) node to the qubits it acts on.
+    nodes: HashMap<NodeIndex, [usize; 2]>,
+    /// Map of each qubit to the node that acts on it and the other qubit that node acts on, if this
+    /// qubit is active (otherwise `None`).
+    qubits: Vec<Option<(NodeIndex, usize)>>,
+    /// Tracking the insertion order of nodes, so iteration can always go through them in a
+    /// deterministic order.  This is important for reproducibility from a set seed - when building
+    /// up the extended set with a fixed, finite size, the iteration order through the nodes of the
+    /// front layer is important.  We need to maintain the insertion order even with removals from
+    /// the layer.
+    iteration_order: Vec<Option<NodeIndex>>,
+    /// The index of the first populated entry in the `iteration_order`.  If the iteration order is
+    /// empty, this will be 0.
+    iteration_start: usize,
+    /// The index one past the last populated entry in the `iteration_order`.  If the iteration
+    /// order is empty, this will be 0.
+    iteration_end: usize,
+}
+
+impl FrontLayer {
+    pub fn new(num_qubits: usize) -> Self {
+        FrontLayer {
+            // This is the maximum capacity of the front layer, since each qubit must be one of a
+            // pair, and can only have one gate in the layer.
+            nodes: HashMap::with_capacity(num_qubits / 2),
+            qubits: vec![None; num_qubits],
+            iteration_order: vec![None; num_qubits],
+            iteration_start: 0,
+            iteration_end: 0,
+        }
+    }
+
+    /// Add a node into the front layer, with the two qubits it operates on.  This usually has
+    /// constant-time complexity, except if the iteration-order buffer is full.
+    pub fn insert(&mut self, index: NodeIndex, qubits: [usize; 2]) {
+        let [a, b] = qubits;
+        self.qubits[a] = Some((index, b));
+        self.qubits[b] = Some((index, a));
+        self.nodes.insert(index, qubits);
+
+        self.iteration_order[self.iteration_end] = Some(index);
+        self.iteration_end += 1;
+        if self.iteration_end == self.iteration_order.len() {
+            // Condense items back to the start of the vector.
+            let mut ptr = 0;
+            for i in self.iteration_start..self.iteration_end {
+                if let Some(value) = self.iteration_order[i] {
+                    self.iteration_order[i] = None;
+                    self.iteration_order[ptr] = Some(value);
+                    ptr += 1;
+                }
+            }
+            self.iteration_start = 0;
+            self.iteration_end = ptr;
+        }
+    }
+
+    /// Remove a node from the front layer.
+    pub fn remove(&mut self, index: &NodeIndex) {
+        let [q0, q1] = self.nodes.remove(index).unwrap();
+        self.qubits[q0] = None;
+        self.qubits[q1] = None;
+
+        // If the element was at the start of the iteration order, advance the pointer.
+        match self.iteration_order[self.iteration_start] {
+            Some(a) if a == *index => {
+                self.iteration_order[self.iteration_start] = None;
+                if self.iteration_start + 1 == self.iteration_end {
+                    self.iteration_start = 0;
+                    self.iteration_end = 0;
+                }
+                while self.iteration_start < self.iteration_end
+                    && self.iteration_order[self.iteration_start].is_none()
+                {
+                    self.iteration_start += 1;
+                }
+            }
+            _ => (),
+        }
+        // Search through and remove the element.  We leave a gap and preserve the insertion order.
+        for i in (self.iteration_start + 1)..self.iteration_end {
+            match self.iteration_order[i] {
+                Some(a) if a == *index => {
+                    self.iteration_order[i] = None;
+                    break;
+                }
+                _ => (),
+            }
+        }
+    }
+
+    /// Query whether a qubit has an active node.
+    #[inline]
+    pub fn is_active(&self, qubit: usize) -> bool {
+        self.qubits[qubit].is_some()
+    }
+
+    /// Calculate the score _difference_ caused by this swap, compared to not making the swap.
+    #[inline]
+    pub fn score(&self, swap: [usize; 2], layout: &NLayout, dist: &ArrayView2<f64>) -> f64 {
+        if self.is_empty() {
+            return 0.0;
+        }
+        // At most there can be two affected gates in the front layer (one on each qubit in the
+        // swap), since any gate whose closest path passes through the swapped qubit link has its
+        // "virtual-qubit path" order changed, but not the total weight.  In theory, we should
+        // never consider the same gate in both `if let` branches, because if we did, the gate would
+        // already be routable.  It doesn't matter, though, because the two distances would be
+        // equal anyway, so not affect the score.
+        let [a, b] = swap;
+        let mut total = 0.0;
+        if let Some((_, c)) = self.qubits[a] {
+            let p_c = layout.logic_to_phys[c];
+            total += dist[[layout.logic_to_phys[b], p_c]] - dist[[layout.logic_to_phys[a], p_c]]
+        }
+        if let Some((_, c)) = self.qubits[b] {
+            let p_c = layout.logic_to_phys[c];
+            total += dist[[layout.logic_to_phys[a], p_c]] - dist[[layout.logic_to_phys[b], p_c]]
+        }
+        total / self.nodes.len() as f64
+    }
+
+    /// Calculate the total absolute of the current front layer on the given layer.
+    pub fn total_score(&self, layout: &NLayout, dist: &ArrayView2<f64>) -> f64 {
+        if self.is_empty() {
+            return 0.0;
+        }
+        self.iter()
+            .map(|(_, &[l_a, l_b])| dist[[layout.logic_to_phys[l_a], layout.logic_to_phys[l_b]]])
+            .sum::<f64>()
+            / self.nodes.len() as f64
+    }
+
+    /// Populate a of nodes that would be routable if the given swap was applied to a layout.  This
+    /// mutates `routable` to avoid heap allocations in the main logic loop.
+    pub fn routable_after(
+        &self,
+        routable: &mut Vec<NodeIndex>,
+        swap: &[usize; 2],
+        layout: &NLayout,
+        coupling: &DiGraph<(), ()>,
+    ) {
+        let [a, b] = *swap;
+        if let Some((node, c)) = self.qubits[a] {
+            if coupling.contains_edge(
+                NodeIndex::new(layout.logic_to_phys[b]),
+                NodeIndex::new(layout.logic_to_phys[c]),
+            ) {
+                routable.push(node);
+            }
+        }
+        if let Some((node, c)) = self.qubits[b] {
+            if coupling.contains_edge(
+                NodeIndex::new(layout.logic_to_phys[a]),
+                NodeIndex::new(layout.logic_to_phys[c]),
+            ) {
+                routable.push(node);
+            }
+        }
+    }
+
+    /// True if there are no nodes in the current layer.
+    #[inline]
+    pub fn is_empty(&self) -> bool {
+        self.nodes.is_empty()
+    }
+
+    /// Iterator over the nodes and the pair of qubits they act on.
+    pub fn iter(&self) -> impl Iterator<Item = (&NodeIndex, &[usize; 2])> {
+        (&self.iteration_order)[self.iteration_start..self.iteration_end]
+            .iter()
+            .filter_map(move |node_opt| node_opt.as_ref().map(|node| (node, &self.nodes[node])))
+    }
+
+    /// Iterator over the nodes.
+    pub fn iter_nodes(&self) -> impl Iterator<Item = &NodeIndex> {
+        (&self.iteration_order)[self.iteration_start..self.iteration_end]
+            .iter()
+            .filter_map(|node_opt| node_opt.as_ref())
+    }
+
+    /// Iterator over the qubits that have active nodes on them.
+    pub fn iter_active(&self) -> impl Iterator<Item = &usize> {
+        (&self.iteration_order)[self.iteration_start..self.iteration_end]
+            .iter()
+            .filter_map(move |node_opt| node_opt.as_ref().map(|node| &self.nodes[node]))
+            .flatten()
+    }
+}
+
+/// This is largely similar to the `FrontLayer` struct, but does not need to track the insertion
+/// order of the nodes, and can have more than one node on each active qubit.  This does not have a
+/// `remove` method (and its data structures aren't optimised for fast removal), since the extended
+/// set is built from scratch each time a new gate is routed.
+pub struct ExtendedSet {
+    nodes: HashMap<NodeIndex, [usize; 2]>,
+    qubits: Vec<Vec<usize>>,
+}
+
+impl ExtendedSet {
+    pub fn new(num_qubits: usize, max_size: usize) -> Self {
+        ExtendedSet {
+            nodes: HashMap::with_capacity(max_size),
+            qubits: vec![Vec::new(); num_qubits],
+        }
+    }
+
+    /// Add a node and its active qubits to the extended set.
+    pub fn insert(&mut self, index: NodeIndex, qubits: &[usize; 2]) -> bool {
+        let [a, b] = *qubits;
+        if self.nodes.insert(index, *qubits).is_none() {
+            self.qubits[a].push(b);
+            self.qubits[b].push(a);
+            true
+        } else {
+            false
+        }
+    }
+
+    /// Calculate the score of applying the given swap, relative to not applying it.
+    pub fn score(&self, swap: [usize; 2], layout: &NLayout, dist: &ArrayView2<f64>) -> f64 {
+        if self.nodes.is_empty() {
+            return 0.0;
+        }
+        let [l_a, l_b] = swap;
+        let p_a = layout.logic_to_phys[l_a];
+        let p_b = layout.logic_to_phys[l_b];
+        let mut total = 0.0;
+        for &l_other in self.qubits[l_a].iter() {
+            // If the other qubit is also active then the score won't have changed, but since the
+            // distance is absolute, we'd double count rather than ignore if we didn't skip it.
+            if l_other == l_b {
+                continue;
+            }
+            let p_other = layout.logic_to_phys[l_other];
+            total += dist[[p_b, p_other]] - dist[[p_a, p_other]];
+        }
+        for &l_other in self.qubits[l_b].iter() {
+            if l_other == l_a {
+                continue;
+            }
+            let p_other = layout.logic_to_phys[l_other];
+            total += dist[[p_a, p_other]] - dist[[p_b, p_other]];
+        }
+        total / self.nodes.len() as f64
+    }
+
+    /// Calculate the total absolute score of this set of nodes over the given layout.
+    pub fn total_score(&self, layout: &NLayout, dist: &ArrayView2<f64>) -> f64 {
+        if self.nodes.is_empty() {
+            return 0.0;
+        }
+        self.nodes
+            .iter()
+            .map(|(_, &[l_a, l_b])| dist[[layout.logic_to_phys[l_a], layout.logic_to_phys[l_b]]])
+            .sum::<f64>()
+            / self.nodes.len() as f64
+    }
+
+    /// Clear all nodes from the extended set.
+    pub fn clear(&mut self) {
+        for &[a, b] in self.nodes.values() {
+            self.qubits[a].clear();
+            self.qubits[b].clear();
+        }
+        self.nodes.clear()
+    }
+
+    /// Number of nodes in the set.
+    pub fn len(&self) -> usize {
+        self.nodes.len()
+    }
+}