Skip to content

Speedup sorting for inline views: 1.4x - 1.7x improvement #7856

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.

Sign up for GitHub

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

Merged
merged 8 commits into from
Jul 3, 2025
Merged

Speedup sorting for inline views: 1.4x - 1.7x improvement #7856

Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
8b74554
speedup byteview comparisons
Dandandan Jul 3, 2025
04f2b1a
revert sort, don't store len
Dandandan Jul 3, 2025
9ed518f
Add all-inline case
Dandandan Jul 3, 2025
92cacf0
Remove cmp changes
Dandandan Jul 3, 2025
97e9265
Cleanup
Dandandan Jul 3, 2025
eea04dc
Cleanup
Dandandan Jul 3, 2025
b2a1dd0
Remove comment
Dandandan Jul 3, 2025
fb7e260
Update arrow-ord/src/sort.rs
Dandandan Jul 3, 2025
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
100 changes: 60 additions & 40 deletions arrow-ord/src/sort.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -311,51 +311,71 @@ fn sort_byte_view<T: ByteViewType>(
limit: Option<usize>,
) -> UInt32Array {
// 1. Build a list of (index, raw_view, length)
let mut valids: Vec<_> = value_indices
.into_iter()
.map(|idx| {
// SAFETY: we know idx < values.len()
let raw = unsafe { *values.views().get_unchecked(idx as usize) };
let len = raw as u32; // lower 32 bits encode length
(idx, raw, len)
})
.collect();

let mut valids: Vec<_>;
// 2. Compute the number of non-null entries to partially sort
let vlimit = match (limit, options.nulls_first) {
(Some(l), true) => l.saturating_sub(nulls.len()).min(valids.len()),
_ => valids.len(),
let vlimit: usize = match (limit, options.nulls_first) {
(Some(l), true) => l.saturating_sub(nulls.len()).min(value_indices.len()),
_ => value_indices.len(),
};
// 3.a Check if all views are inline (no data buffers)
if values.data_buffers().is_empty() {
Copy link
Contributor

Choose a reason for hiding this comment

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

Thank you @Dandandan , it's very clear we optimize the no data buffers path!

valids = value_indices
.into_iter()
.map(|idx| {
// SAFETY: we know idx < values.len()
let raw = unsafe { *values.views().get_unchecked(idx as usize) };
let inline_key = GenericByteViewArray::<T>::inline_key_fast(raw);
(idx, inline_key)
})
.collect();
let cmp_inline = |a: &(u32, u128), b: &(u32, u128)| a.1.cmp(&b.1);

// 3. Mixed comparator: first prefix, then inline vs full comparison
let cmp_mixed = |a: &(u32, u128, u32), b: &(u32, u128, u32)| {
let (_, raw_a, len_a) = *a;
let (_, raw_b, len_b) = *b;

// 3.1 Both inline (≤12 bytes): compare full 128-bit key including length
if len_a <= MAX_INLINE_VIEW_LEN && len_b <= MAX_INLINE_VIEW_LEN {
return GenericByteViewArray::<T>::inline_key_fast(raw_a)
.cmp(&GenericByteViewArray::<T>::inline_key_fast(raw_b));
// Partially sort according to ascending/descending
if !options.descending {
sort_unstable_by(&mut valids, vlimit, cmp_inline);
} else {
sort_unstable_by(&mut valids, vlimit, |x, y| cmp_inline(x, y).reverse());
}
} else {
valids = value_indices
.into_iter()
.map(|idx| {
// SAFETY: we know idx < values.len()
let raw = unsafe { *values.views().get_unchecked(idx as usize) };
(idx, raw)
})
.collect();
// 3.b Mixed comparator: first prefix, then inline vs full comparison
let cmp_mixed = |a: &(u32, u128), b: &(u32, u128)| {
let (_, raw_a) = *a;
let (_, raw_b) = *b;
let len_a = raw_a as u32;
let len_b = raw_b as u32;
// 3.b.1 Both inline (≤12 bytes): compare full 128-bit key including length
if len_a <= MAX_INLINE_VIEW_LEN && len_b <= MAX_INLINE_VIEW_LEN {
return GenericByteViewArray::<T>::inline_key_fast(raw_a)
.cmp(&GenericByteViewArray::<T>::inline_key_fast(raw_b));
}

// 3.2 Compare 4-byte prefix in big-endian order
let pref_a = ByteView::from(raw_a).prefix.swap_bytes();
let pref_b = ByteView::from(raw_b).prefix.swap_bytes();
if pref_a != pref_b {
return pref_a.cmp(&pref_b);
}
// 3.b.2 Compare 4-byte prefix in big-endian order
let pref_a = ByteView::from(raw_a).prefix.swap_bytes();
let pref_b = ByteView::from(raw_b).prefix.swap_bytes();
if pref_a != pref_b {
return pref_a.cmp(&pref_b);
}

// 3.3 Fallback to full byte-slice comparison
let full_a: &[u8] = unsafe { values.value_unchecked(a.0 as usize).as_ref() };
let full_b: &[u8] = unsafe { values.value_unchecked(b.0 as usize).as_ref() };
full_a.cmp(full_b)
};
// 3.b.3 Fallback to full byte-slice comparison
let full_a: &[u8] = unsafe { values.value_unchecked(a.0 as usize).as_ref() };
let full_b: &[u8] = unsafe { values.value_unchecked(b.0 as usize).as_ref() };
full_a.cmp(full_b)
};

// 4. Partially sort according to ascending/descending
if !options.descending {
sort_unstable_by(&mut valids, vlimit, cmp_mixed);
} else {
sort_unstable_by(&mut valids, vlimit, |x, y| cmp_mixed(x, y).reverse());
// 3.b.4 Partially sort according to ascending/descending
if !options.descending {
sort_unstable_by(&mut valids, vlimit, cmp_mixed);
} else {
sort_unstable_by(&mut valids, vlimit, |x, y| cmp_mixed(x, y).reverse());
}
}

// 5. Assemble nulls and sorted indices into final output
Expand All @@ -367,10 +387,10 @@ fn sort_byte_view<T: ByteViewType>(
// Place null indices first
out.extend_from_slice(&nulls[..nulls.len().min(out_limit)]);
let rem = out_limit - out.len();
out.extend(valids.iter().map(|&(i, _, _)| i).take(rem));
out.extend(valids.iter().map(|&(i, _)| i).take(rem));
} else {
// Place non-null indices first
out.extend(valids.iter().map(|&(i, _, _)| i).take(out_limit));
out.extend(valids.iter().map(|&(i, _)| i).take(out_limit));
let rem = out_limit - out.len();
out.extend_from_slice(&nulls[..rem]);
}
Expand Down
Loading