rcb: don't split small item sets in parallel

It's slower to go through rayon than to sort+split them.

Author: hhirtz

src/algorithms/recursive_bisection.rs

@@ -26,28 +26,86 @@ struct Item<'p, const D: usize, W> {
     part: &'p AtomicUsize,
 }
 
-/// Return value of [rcb_split].
-struct SplitResult<W> {
-    /// Index of the first item in the right part in the array of [Item]s.
-    split_idx: usize,
+/// Return value of [rcb_split] and [par_rcb_split].
+struct SplitResult<'a, 'p, const D: usize, W> {
+    left: &'a mut [Item<'p, D, W>],
+    right: &'a mut [Item<'p, D, W>],
     /// Weight of the left part, used to compute the sum for the next iteration.
     weight_left: W,
     /// Coordinate value of the split, used to compute the [BoundingBox]es.
     split_pos: f64,
 }
 
-fn rcb_split<const D: usize, W>(
-    items: &mut [Item<D, W>],
+/// Splits the given items into two sets of similar weights.
+fn rcb_split<'a, 'p, const D: usize, W>(
+    items: &'a mut [Item<'p, D, W>],
+    coord: usize,
+    max: f64,
+    sum: W,
+) -> SplitResult<'a, 'p, D, W>
+where
+    W: RcbWeight,
+{
+    let span = tracing::info_span!("rcb_split");
+    let _enter = span.enter();
+
+    items.sort_unstable_by(|item1, item2| {
+        crate::partial_cmp(&item1.point[coord], &item2.point[coord])
+    });
+
+    let mut count_left = items.len();
+    let mut split_pos = max;
+    let mut weight_left = W::default();
+    for (i, item) in items.iter().enumerate() {
+        if sum <= weight_left + weight_left {
+            count_left = i;
+            split_pos = item.point[coord];
+            break;
+        }
+        weight_left += item.weight;
+    }
+
+    let (left, right) = items.split_at_mut(count_left);
+
+    SplitResult {
+        left,
+        right,
+        weight_left,
+        split_pos,
+    }
+}
+
+/// Wrapper around [`slice::split_at_mut`] which makes it so elements of the
+/// left side are all lower than all elements of the right side.
+fn reorder_split<F, T>(s: &mut [T], index: usize, compare: F) -> (&mut [T], &mut [T])
+where
+    F: Fn(&T, &T) -> cmp::Ordering,
+{
+    if index == s.len() {
+        s.split_at_mut(s.len())
+    } else {
+        let span = tracing::info_span!("select_nth_unstable");
+        let _enter = span.enter();
+
+        let (left, _, _right_minus_one) = s.select_nth_unstable_by(index, compare);
+        let left_len = left.len();
+        s.split_at_mut(left_len)
+    }
+}
+
+/// Splits the given items into two sets of similar weights (parallel version).
+fn par_rcb_split<'a, 'p, const D: usize, W>(
+    items: &'a mut [Item<'p, D, W>],
     coord: usize,
     tolerance: f64,
     mut min: f64,
     mut max: f64,
     sum: W,
-) -> SplitResult<W>
+) -> SplitResult<'a, 'p, D, W>
 where
     W: RcbWeight,
 {
-    let span = tracing::info_span!("rcb_split");
+    let span = tracing::info_span!("par_rcb_split");
     let _enter = span.enter();
 
     let mut prev_count_left = usize::MAX;
@@ -71,8 +129,12 @@ where
             f64::abs((weight_left - ideal_weight_left) / ideal_weight_left)
         };
         if count_left == prev_count_left || imbalance < tolerance {
+            let (left, right) = reorder_split(items, count_left, |item1, item2| {
+                crate::partial_cmp(&item1.point[coord], &item2.point[coord])
+            });
             return SplitResult {
-                split_idx: count_left,
+                left,
+                right,
                 weight_left,
                 split_pos: split_target,
             };
@@ -117,37 +179,24 @@ fn rcb_recurse<const D: usize, W>(
         return;
     }
 
-    let span = tracing::info_span!("rcb_recurse");
-    let enter = span.enter();
-
     let min = bb.p_min[coord];
     let max = bb.p_max[coord];
     let SplitResult {
-        split_idx,
+        left,
+        right,
         weight_left,
         split_pos,
-    } = rcb_split(items, coord, tolerance, min, max, sum);
-    let (left, right) = if split_idx == items.len() {
-        items.split_at_mut(items.len())
+    } = if items.len() > 4096 {
+        par_rcb_split(items, coord, tolerance, min, max, sum)
     } else {
-        let span = tracing::info_span!("select_nth_unstable");
-        let _enter = span.enter();
-
-        let (left, _, _right_minus_one) = items
-            .select_nth_unstable_by(split_idx, |item1, item2| {
-                crate::partial_cmp(&item1.point[coord], &item2.point[coord])
-            });
-        let left_len = left.len();
-        items.split_at_mut(left_len)
+        rcb_split(items, coord, max, sum)
     };
 
     let mut bb_left = bb.clone();
     bb_left.p_max[coord] = split_pos;
     let mut bb_right = bb;
     bb_right.p_min[coord] = split_pos;
 
-    mem::drop(enter);
-
     rayon::join(
         || {
             rcb_recurse(