Repository navigation

Blocked `BitPacked`: fuse `FoR` decode over blocked children - #10251

Draft

mhk197 wants to merge 27 commits into

mk/bitpacked-v2-schemefrom

mk/for-bitpacked-blocked-fused

Commits on Oct 2, 2026
- store bitpacked block offsets as an optional child
  mhk197
  committed
  abf3355
- check bitpacked widths once per kernel
  mhk197
  committed
  477bfb9
- only check bitpacked block boundaries in debug builds
  mhk197
  committed
  5f3f2c3
- move bitpack_compress into a directory
  mhk197
  committed
  cf2c375
- encode bitpacked arrays with a width per block
  mhk197
  committed
  861873b
- rename bitpacked encode benches
  mhk197
  committed
  1729b6d
- only zero-pad a partial last block when bitpacking
  mhk197
  committed
  b7bcace
- keep the global bitpack encoder unchanged
  mhk197
  committed
  b369fb6
- move global-only bitpacking code back into global.rs
  mhk197
  committed
  8c525af
- fix clippy
  mhk197
  committed
  a109b8c
- bench only the blocked packing kernel
  mhk197
  committed
  95f5480
- fix comment
  mhk197
  committed
  de28cfc
- move bitpack_decompress into a directory
  mhk197
  committed
  0d407fd
- decode bitpacked blocks with variable bit widths
  mhk197
  committed
  385b15a
- trim blocked decode bench to one case per type
  mhk197
  committed
  c8d26f8
- round-trip blocked bitpacking and encode the decode bench input
  mhk197
  committed
  67da367
- dispatch bitpacked decode in callers and drop blocked mapped decode
  mhk197
  committed
  16b8075
- serialize blocked bitpacked arrays as fastlanes.bitpacked.v2
  mhk197
  committed
  94c7da6
- fix import order
  mhk197
  committed
  ca053fd
- refine BitPackingScheme to per-block bit widths when fastlanes.bitpacked.v2 is allowed
  mhk197
  committed
  c61995a
- fix clippy
  mhk197
  committed
  336bbfc
- Merge branch 'mk/bitpacked-v2-wire' into mk/bitpacked-v2-scheme
  mhk197
  committed
  a500896
- always produce blocked arrays in bitpacking v2
  mhk197
  committed
  9b4777b
- keep the dictionary test out of blocked bitpacking's reach
  mhk197
  committed
  c98cb99
- fuse FoR decode over blocked bitpacked children
  mhk197
  committed
  1de55bc
- widen bitpacked block offsets without a cast
  mhk197
  committed
  a7707ed
- bench FoR decode over a blocked bitpacked child
  mhk197
  committed
  0b18f5c

Navigation Menu

Repository navigation

Blocked `BitPacked`: fuse `FoR` decode over blocked children - #10251

Blocked `BitPacked`: fuse `FoR` decode over blocked children#10251
mhk197 wants to merge 27 commits into
mk/bitpacked-v2-schemefrom
mk/for-bitpacked-blocked-fused

Commits on Oct 2, 2026

store bitpacked block offsets as an optional child

check bitpacked widths once per kernel

only check bitpacked block boundaries in debug builds

move bitpack_compress into a directory

encode bitpacked arrays with a width per block

rename bitpacked encode benches

only zero-pad a partial last block when bitpacking

keep the global bitpack encoder unchanged

move global-only bitpacking code back into global.rs

fix clippy

bench only the blocked packing kernel

fix comment

move bitpack_decompress into a directory

decode bitpacked blocks with variable bit widths

trim blocked decode bench to one case per type

round-trip blocked bitpacking and encode the decode bench input

dispatch bitpacked decode in callers and drop blocked mapped decode

serialize blocked bitpacked arrays as fastlanes.bitpacked.v2

fix import order

refine BitPackingScheme to per-block bit widths when fastlanes.bitpacked.v2 is allowed

fix clippy

Merge branch 'mk/bitpacked-v2-wire' into mk/bitpacked-v2-scheme

always produce blocked arrays in bitpacking v2

keep the dictionary test out of blocked bitpacking's reach

fuse FoR decode over blocked bitpacked children

widen bitpacked block offsets without a cast

bench FoR decode over a blocked bitpacked child