replace builder with bulk take in naive_rebuild

Signed-off-by: Baris Palaska <barispalaska@gmail.com>

Author: palaska

vortex-array/Cargo.toml

@@ -169,5 +169,9 @@ harness = false
 name = "take_fsl"
 harness = false
 
+[[bench]]
+name = "listview_rebuild"
+harness = false
+
 [package.metadata.cargo-machete]
 ignored = ["getrandom_v03"]

vortex-array/benches/listview_rebuild.rs

@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright the Vortex contributors
+
+//! Benchmarks for ListView rebuild with primitive and VarBinView elements.
+
+#![allow(clippy::unwrap_used)]
+#![allow(clippy::cast_possible_truncation)]
+
+use divan::Bencher;
+use vortex_array::IntoArray;
+use vortex_array::arrays::ListViewArray;
+use vortex_array::arrays::ListViewRebuildMode;
+use vortex_array::arrays::PrimitiveArray;
+use vortex_array::arrays::VarBinViewArray;
+use vortex_array::validity::Validity;
+use vortex_buffer::Buffer;
+
+fn main() {
+    divan::main();
+}
+
+const NUM_LISTS: &[usize] = &[100, 1_000, 10_000];
+
+/// Build a ListViewArray of primitives with overlapping views.
+fn make_primitive_listview(num_lists: usize) -> ListViewArray {
+    let element_count = num_lists * 4;
+    let elements = PrimitiveArray::from_iter(0..element_count as i32).into_array();
+
+    let offsets: Buffer<u32> = (0..num_lists).map(|i| (i * 2) as u32).collect();
+    let sizes: Buffer<u32> = std::iter::repeat_n(4u32, num_lists).collect();
+
+    ListViewArray::new(
+        elements,
+        offsets.into_array(),
+        sizes.into_array(),
+        Validity::NonNullable,
+    )
+}
+
+/// Build a ListViewArray of strings with overlapping views.
+fn make_varbinview_listview(num_lists: usize) -> ListViewArray {
+    let element_count = num_lists * 4;
+    let strings: Vec<String> = (0..element_count)
+        .map(|i| {
+            if i % 3 == 0 {
+                format!("long-string-value-{i:06}")
+            } else {
+                format!("s{i}")
+            }
+        })
+        .collect();
+    let elements = VarBinViewArray::from_iter_str(strings.iter().map(|s| s.as_str())).into_array();
+
+    let offsets: Buffer<u32> = (0..num_lists).map(|i| (i * 2) as u32).collect();
+    let sizes: Buffer<u32> = std::iter::repeat_n(4u32, num_lists).collect();
+
+    ListViewArray::new(
+        elements,
+        offsets.into_array(),
+        sizes.into_array(),
+        Validity::NonNullable,
+    )
+}
+
+#[divan::bench(args = NUM_LISTS)]
+fn rebuild_primitive(bencher: Bencher, num_lists: usize) {
+    let listview = make_primitive_listview(num_lists);
+    bencher
+        .with_inputs(|| &listview)
+        .bench_refs(|lv| lv.rebuild(ListViewRebuildMode::MakeZeroCopyToList).unwrap());
+}
+
+#[divan::bench(args = NUM_LISTS)]
+fn rebuild_varbinview(bencher: Bencher, num_lists: usize) {
+    let listview = make_varbinview_listview(num_lists);
+    bencher
+        .with_inputs(|| &listview)
+        .bench_refs(|lv| lv.rebuild(ListViewRebuildMode::MakeZeroCopyToList).unwrap());
+}

vortex-array/src/arrays/listview/rebuild.rs

@@ -14,7 +14,6 @@ use crate::Array;
 use crate::IntoArray;
 use crate::ToCanonical;
 use crate::arrays::ListViewArray;
-use crate::builders::builder_with_capacity;
 use crate::compute;
 use crate::vtable::ValidityHelper;
 
@@ -103,18 +102,14 @@ impl ListViewArray {
         })
     }
 
-    // TODO(connor)[ListView]: We should benchmark if it is faster to use `take` on the elements
-    // instead of using a builder.
     /// The inner function for `rebuild_zero_copy_to_list`, which rebuilds a `ListViewArray` piece
     /// by piece.
+    ///
+    /// Collects all element indices into a flat `BufferMut<u64>` and performs a single bulk
+    /// `take` to produce the new elements array.
     fn naive_rebuild<O: IntegerPType, NewOffset: IntegerPType, S: IntegerPType>(
         &self,
     ) -> VortexResult<ListViewArray> {
-        let element_dtype = self
-            .dtype()
-            .as_list_element_opt()
-            .vortex_expect("somehow had a canonical list that was not a list");
-
         let offsets_canonical = self.offsets().to_primitive();
         let offsets_slice = offsets_canonical.as_slice::<O>();
         let sizes_canonical = self.sizes().to_primitive();
@@ -129,17 +124,8 @@ impl ListViewArray {
         // null lists to 0.
         let mut new_sizes = BufferMut::<S>::with_capacity(len);
 
-        // Canonicalize the elements up front as we will be slicing the elements quite a lot.
-        let elements_canonical = self
-            .elements()
-            .to_canonical()
-            .vortex_expect("canonicalize elements for rebuild")
-            .into_array();
-
-        // Note that we do not know what the exact capacity should be of the new elements since
-        // there could be overlaps in the existing `ListViewArray`.
-        let mut new_elements_builder =
-            builder_with_capacity(element_dtype.as_ref(), self.elements().len());
+        // Collect all element indices for a single bulk take.
+        let mut take_indices = BufferMut::<u64>::with_capacity(self.elements().len());
 
         let mut n_elements = NewOffset::zero();
         for index in 0..len {
@@ -159,14 +145,14 @@ impl ListViewArray {
 
             new_offsets.push(n_elements);
             new_sizes.push(size);
-            new_elements_builder.extend_from_array(&elements_canonical.slice(start..stop)?);
+            take_indices.extend(start as u64..stop as u64);
 
             n_elements += num_traits::cast(size).vortex_expect("Cast failed");
         }
 
         let offsets = new_offsets.into_array();
         let sizes = new_sizes.into_array();
-        let elements = new_elements_builder.finish();
+        let elements = self.elements().take(take_indices.into_array())?;
 
         debug_assert_eq!(
             n_elements.as_(),