Add CompactFor encoding — claiming EncodingType::CompactFor = 17 (#791)#791
Open
duxiao1212 wants to merge 3 commits into
Open
Add CompactFor encoding — claiming EncodingType::CompactFor = 17 (#791)#791duxiao1212 wants to merge 3 commits into
duxiao1212 wants to merge 3 commits into
Conversation
meta-cla
Bot
added
the
CLA Signed
|
@duxiao1212 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D105755627. |
duxiao1212
added a commit
to duxiao1212/nimble
that referenced
this pull request
May 28, 2026
…ebookincubator#791) Summary: Pull Request resolved: facebookincubator#791 Space-efficient FOR encoder specialised for `int64`/`uint64` (slot 17). **Relationship to `FixedBitWidth`** Same core algorithm (subtract baseline, bitpack residuals via `FixedBitArray`), but two wire-format improvements: | | `FixedBitWidth` | `CompactFor` | Savings | |---|---|---|---| | **Baseline** | Fixed `sizeof(T)` (8B for int64) | Varint (1-10B) | Up to 7B for int64; diminishes for smaller types (int32: max 3B, int8: ~0) | | **BitWidth** | Rounded up to byte boundary (4→8) | Exact value (4 stays 4) | Depends on data; can halve the bitpacked region | Everything else — `FixedBitArray::bulkGet64WithBaseline` decode, `FixedBitArray::bufferSize` sizing, `FixedBitArray::set`/`bulkSet32` encode — is identical shared infrastructure. No code is duplicated; only the 10-line wire-format header differs. **Why int64 only (not all integer types)** The varint baseline savings scale with `sizeof(T)`. For int64 (8B → 1-10B varint) the savings are substantial. For int32 (4B → 1-5B) they are moderate, and Pfor/SimdForBitpack already dominate those types in benchmarks (13/19 trait wins). For int8/int16 (1-2B) there is no savings at all. Extending to smaller types is trivial (remove `static_assert`, widen factory dispatch) but produces no practical benefit. **Wire format** (after the standard 6-byte Encoding prefix): ``` [varint: baseline] — min value (1-10 bytes) [1 byte: bitWidth] — bits per residual (0-64, exact) [bitpacked residuals] — FixedBitArray::bufferSize(N, bitWidth) bytes (omitted when bitWidth == 0) ``` **Encode**: When `bitWidth <= 32` (common — most int64 ranges fit in 32 bits), uses the template-unrolled `FixedBitArray::bulkSet32` fast path. Falls back to per-element `set()` for wider bit widths. **Decode**: `FixedBitArray::bulkGet64WithBaseline(row, count, output, baseline)` — the same branchless bulk path used by `FixedBitWidthEncoding`, with a single-load fast path for bitWidth ≤ 56. **readWithVisitor**: Implements `bulkScan()` for the `readWithVisitorFast` path (unlike Pfor which uses the slow path). O(1) random access — no sequential state. **Auto-selection**: Excluded from `possibleEncodings()` and `defaultReadFactors()`. Available only via explicit `EncodingLayoutTree` overrides until production bake-in. Ported from AUS `LONG_LIST_FOR_BITPACKED`. Differential Revision: D105755627
meta-codesync
Bot
changed the title
duxiao1212
force-pushed
the
export-D105755627
branch
from
859da83 to
6c2ddf0
Compare
Summary: PFOR is a sparse-outlier patched Frame-of-Reference encoder ported from MRS's AusList SOTA encoder portfolio (D97646777). Bitpacks the bottom 90th-percentile of residuals at base bitwidth, stores the top 10% as (position, full value) exception patches. Wins on columns with rare large outliers where a pure FOR would inflate bitwidth to cover the worst case. Wire format: [encType:1B][dataType:1B][rowCount:4B][min:GV32][baseBitWidth:1B][numExceptions:varint][exceptionPositions:varint[]][exceptionValues:varint[]][bitpacked baseResiduals][7B zero-pad tail] Decode follows the post-D98819389-style optimization: pass 1 bitunpacks all residuals via byte-aligned loadU64; pass 2 patches the exceptions in. Slot allocation: claiming EncodingType::Pfor = 15. Per WS8 audit (T271476729), slots 12/13/14 are contested by 4+ in-flight diffs (D103045783 ChunkedBitPacking=12, D103975634 FOR=13/SubIntSplit=14, D105064770/D105134978 Fsst=12/14, plus implied ChunkedALP=13). 15 is the next free slot at time of commit. If a contending diff lands first claiming 15, this diff will need rebase to next free. Initial readFactor: 1.5 (conservative). Prevents auto-selection by ManualEncodingSelectionPolicy until baked in. Mirrors ChunkedBitPacking pattern. Stale-binary risk: xldb/disco/user_experiments/util/WriteConfigurationUtil.cpp persists EncodingType through Configerator. Older readers will throw NIMBLE_UNREACHABLE on EncodingType::Pfor=15 until they pick up this build — fail-loud, not silent. Differential Revision: D105270894
Summary: Second-order delta encoder for int64 columns: take double-deltas, zigzag-encode, FOR-baseline-shift, bitpack at minimum bitwidth. Wins on regular-interval timestamp data. Renamed from LongDoubleDeltaBitpack for naming consistency (no type prefix in Nimble convention). Differential Revision: D105357210
…ebookincubator#791) Summary: Pull Request resolved: facebookincubator#791 Space-efficient FOR encoder specialised for `int64`/`uint64` (slot 17). **Relationship to `FixedBitWidth`** Same core algorithm (subtract baseline, bitpack residuals via `FixedBitArray`), but two wire-format improvements: | | `FixedBitWidth` | `CompactFor` | Savings | |---|---|---|---| | **Baseline** | Fixed `sizeof(T)` (8B for int64) | Varint (1-10B) | Up to 7B for int64; diminishes for smaller types (int32: max 3B, int8: ~0) | | **BitWidth** | Rounded up to byte boundary (4→8) | Exact value (4 stays 4) | Depends on data; can halve the bitpacked region | Everything else — `FixedBitArray::bulkGet64WithBaseline` decode, `FixedBitArray::bufferSize` sizing, `FixedBitArray::set`/`bulkSet32` encode — is identical shared infrastructure. No code is duplicated; only the 10-line wire-format header differs. **Why int64 only (not all integer types)** The varint baseline savings scale with `sizeof(T)`. For int64 (8B → 1-10B varint) the savings are substantial. For int32 (4B → 1-5B) they are moderate, and Pfor/SimdForBitpack already dominate those types in benchmarks (13/19 trait wins). For int8/int16 (1-2B) there is no savings at all. Extending to smaller types is trivial (remove `static_assert`, widen factory dispatch) but produces no practical benefit. **Wire format** (after the standard 6-byte Encoding prefix): ``` [varint: baseline] — min value (1-10 bytes) [1 byte: bitWidth] — bits per residual (0-64, exact) [bitpacked residuals] — FixedBitArray::bufferSize(N, bitWidth) bytes (omitted when bitWidth == 0) ``` **Encode**: When `bitWidth <= 32` (common — most int64 ranges fit in 32 bits), uses the template-unrolled `FixedBitArray::bulkSet32` fast path. Falls back to per-element `set()` for wider bit widths. **Decode**: `FixedBitArray::bulkGet64WithBaseline(row, count, output, baseline)` — the same branchless bulk path used by `FixedBitWidthEncoding`, with a single-load fast path for bitWidth ≤ 56. **readWithVisitor**: Implements `bulkScan()` for the `readWithVisitorFast` path (unlike Pfor which uses the slow path). O(1) random access — no sequential state. **Auto-selection**: Excluded from `possibleEncodings()` and `defaultReadFactors()`. Available only via explicit `EncodingLayoutTree` overrides until production bake-in. Ported from AUS `LONG_LIST_FOR_BITPACKED`. Differential Revision: D105755627
duxiao1212
force-pushed
the
export-D105755627
branch
from
6c2ddf0 to
71fd3f3
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
Space-efficient FOR encoder specialised for
int64/uint64(slot 17).Relationship to
FixedBitWidthSame core algorithm (subtract baseline, bitpack residuals via
FixedBitArray), but two wire-format improvements:FixedBitWidthCompactForsizeof(T)(8B for int64)Everything else —
FixedBitArray::bulkGet64WithBaselinedecode,FixedBitArray::bufferSizesizing,FixedBitArray::set/bulkSet32encode — is identical shared infrastructure. No code is duplicated; only the 10-line wire-format header differs.Why int64 only (not all integer types)
The varint baseline savings scale with
sizeof(T). For int64 (8B → 1-10B varint) the savings are substantial. For int32 (4B → 1-5B) they are moderate, and Pfor/SimdForBitpack already dominate those types in benchmarks (13/19 trait wins). For int8/int16 (1-2B) there is no savings at all. Extending to smaller types is trivial (removestatic_assert, widen factory dispatch) but produces no practical benefit.Wire format (after the standard 6-byte Encoding prefix):
Encode: When
bitWidth <= 32(common — most int64 ranges fit in 32 bits), uses the template-unrolledFixedBitArray::bulkSet32fast path. Falls back to per-elementset()for wider bit widths.Decode:
FixedBitArray::bulkGet64WithBaseline(row, count, output, baseline)— the same branchless bulk path used byFixedBitWidthEncoding, with a single-load fast path for bitWidth ≤ 56.readWithVisitor: Implements
bulkScan()for thereadWithVisitorFastpath (unlike Pfor which uses the slow path). O(1) random access — no sequential state.Auto-selection: Excluded from
possibleEncodings()anddefaultReadFactors(). Available only via explicitEncodingLayoutTreeoverrides until production bake-in.Ported from AUS
LONG_LIST_FOR_BITPACKED.Differential Revision: D105755627