Skip to content

Fast FastLanes Delta: fused decode, SIMD transposes, probe, filter, take and is_constant kernels - #11428

Merged
robert3005 merged 11 commits into
developfrom
claude/stoic-brahmagupta-p44si2
Nov 13, 2026
Merged

robert3005 merged 11 commits into
developfrom
claude/stoic-brahmagupta-p44si2

Conversation

@robert3005

@robert3005 robert3005 commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

Summary

FastLanes Delta decoded slowly and had almost no compute kernels: every filter, take, scalar_at or is_constant decoded the whole array first. Vortex now has the APIs this needs (the probe API and execute kernels). This PR uses them, together with fastlanes 0.8.0's fused unpack-and-undelta and SIMD transpose kernels, so Delta is cheap enough to select by default.

Changes

Decode

  • Fused decode. Delta(BitPacked) decodes each 1024-value chunk with fastlanes' fused unchecked_undelta_pack. The deltas are no longer unpacked into a full intermediate array. This covers patched and block-width bit-packing.
  • Sliced arrays decode only the chunks the slice covers, then trim the result.
  • SIMD transposes. fastlanes is updated to 0.8.0. Transposes are dispatched at runtime through fearless_simd: AVX-512 for 8- and 16-bit values; for 32- and 64-bit values, fastlanes' own dispatch, because AVX-512 is slower there on Zen 4.
  • DeltaChunks is a reusable chunk decoder with scratch space. The kernels below share it.

Kernels

  • Probe API. scalar_at decodes only the chunk holding the row. A repeated probe caches the last decoded chunk.
  • Filter and take decode only the chunks that have selected rows. Take falls back to a full decode for large index sets.
  • is_constant reads only the stored parts where it can: all-null, mixed-null, or all-zero interior deltas plus constant bases. Only the partial chunks at either end of a slice are decoded.
  • The chunked take logic in take_chunks.rs is shared between BitPacked and Delta.

Compression

  • Back-fill leading nulls per chunk. delta_compress used to carry the previous chunk's last value into a chunk that starts with nulls. Values that restart in every chunk, such as RLE indices, then got wrapping residuals, and the compressor rejected Delta for them. Those leading nulls now take the chunk's first valid value.
  • Keep RunEnd off RLE indices. RLE indices count up within each chunk, so Delta packs them into a few bits. Allowing RunEnd there produced RLE(indices = RunEnd(ends = Delta)), which every decode and filter expands twice. This enables the exclusion that had been commented out.

Loading
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

changelog/feature A new feature

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants