Better fixed-width buffer filtering - #8723
Open
robert3005 wants to merge 7 commits into
Open
Conversation
Merging this PR will regress 2 benchmarks
Warning Please fix the performance issues or acknowledge them on CodSpeed. Performance Changes
Tip Investigate this regression by commenting Comparing 
Footnotes
|
Contributor
Author
|
Moved benchmarks here #8727 |
robert3005
force-pushed
the
rk/filterbuffer
branch
from
July 13, 2026 08:58
21c1bd7 to
423cee5
Compare
robert3005
force-pushed
the
rk/filterbuffer
branch
from
July 20, 2026 12:32
8633f26 to
5efe3a9
Compare
Contributor
Polar Signals Profiling ResultsLatest Run
Previous Runs (3)
Powered by Polar Signals Cloud |
robert3005
added
the
action/benchmark-sql
github-actions
Bot
removed
the
action/benchmark-sql
Contributor
Benchmarks: Vortex queries 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.028x ➖, 0↑ 0↓)
datafusion / parquet (0.934x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed (0.953x ➖, 0↑ 0↓)
duckdb / parquet (0.993x ➖, 0↑ 0↓)
No file size changes detected. |
Contributor
Benchmarks: PolarSignals Profiling 📖Vortex (geomean): 1.038x ➖ How to read Verdict and Engines
datafusion / vortex-file-compressed (1.038x ➖, 0↑ 3↓)
No file size changes detected. |
Contributor
Benchmarks: FineWeb NVMe 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.964x ➖, 2↑ 0↓)
datafusion / parquet (0.980x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed (0.966x ➖, 2↑ 0↓)
duckdb / parquet (0.996x ➖, 0↑ 0↓)
File Size Changes (2 files changed, -46.3% overall, 0↑ 2↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=1 on NVME 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.997x ➖, 0↑ 0↓)
datafusion / parquet (1.004x ➖, 0↑ 0↓)
datafusion / arrow (1.002x ➖, 1↑ 1↓)
duckdb / vortex-file-compressed (0.989x ➖, 0↑ 0↓)
duckdb / parquet (1.015x ➖, 0↑ 2↓)
File Size Changes (10 files changed, -43.9% overall, 0↑ 10↓)
Totals:
|
Contributor
Benchmarks: TPC-DS SF=1 on NVME 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.978x ➖, 13↑ 3↓)
datafusion / parquet (0.965x ➖, 9↑ 3↓)
duckdb / vortex-file-compressed (0.963x ➖, 4↑ 1↓)
duckdb / parquet (0.997x ➖, 1↑ 1↓)
File Size Changes (25 files changed, -43.5% overall, 0↑ 25↓)
Totals:
|
Contributor
Benchmarks: Clickbench on NVME 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.973x ➖, 3↑ 0↓)
datafusion / parquet (0.997x ➖, 0↑ 1↓)
duckdb / vortex-file-compressed (0.992x ➖, 3↑ 0↓)
duckdb / parquet (1.000x ➖, 0↑ 0↓)
File Size Changes (101 files changed, -39.2% overall, 0↑ 101↓)
Totals:
|
Contributor
Benchmarks: FineWeb S3 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (1.149x ➖, 0↑ 1↓)
datafusion / parquet (1.036x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed (1.148x ➖, 0↑ 0↓)
duckdb / parquet (1.026x ➖, 0↑ 0↓)
|
Contributor
Benchmarks: Statistical and Population Genetics 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
duckdb / vortex-file-compressed (0.985x ➖, 0↑ 0↓)
duckdb / parquet (1.021x ➖, 0↑ 0↓)
File Size Changes (2 files changed, -32.3% overall, 0↑ 2↓)
Totals:
|
Contributor
Benchmarks: Clickbench Sorted on NVME 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.952x ➖, 1↑ 0↓)
datafusion / parquet (0.992x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed (0.996x ➖, 0↑ 2↓)
duckdb / parquet (1.010x ➖, 0↑ 0↓)
File Size Changes (201 files changed, -42.8% overall, 61↑ 140↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=10 on NVME 📖Verdict: No clear signal (environment too noisy confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.952x ➖, 3↑ 1↓)
datafusion / parquet (0.939x ➖, 7↑ 2↓)
datafusion / arrow (0.971x ➖, 5↑ 4↓)
duckdb / vortex-file-compressed (0.957x ➖, 4↑ 0↓)
duckdb / parquet (1.009x ➖, 0↑ 0↓)
File Size Changes (25 files changed, -44.0% overall, 0↑ 25↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=1 on S3 📖Verdict: No clear signal (environment too noisy confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed (0.967x ➖, 1↑ 0↓)
datafusion / parquet (1.050x ➖, 0↑ 1↓)
duckdb / vortex-file-compressed (1.014x ➖, 0↑ 0↓)
duckdb / parquet (1.017x ➖, 0↑ 0↓)
|
robert3005
commented
Jul 20, 2026
robert3005
commented
Jul 20, 2026
Signed-off-by: Robert Kruszewski <github@robertk.io>
Signed-off-by: Robert Kruszewski <github@robertk.io>
Adds a NEON tier to the fixed-width filter compress kernels and retunes the
x86 tier against real AVX-512 hardware.
NEON has no `vpcompress`, but `tbl` *is* a cross-lane byte shuffle, so a
byte-index LUT covering one vector of lanes serves every element width -
unlike AVX2, which has no byte shuffle and so leaves 1/2-byte elements on the
scalar byte-LUT. Lanes per shuffle is 8/8/4/2 for 1/2/4/8-byte elements.
The chunk loop is deliberately branchless. An `m != 0` guard looks like a
cheap skip but is a large pessimization: the mask bits of one chunk are as
random as the data, so `P(m == 0) = (1 - density)^lanes` crosses 0.5 inside
the useful density range for every width and the branch mispredicts
near-maximally right where it would pay off. Guarding cost 1.95x at density
0.5 for 4-byte elements and 1.45x at 0.1 for 1-byte. Storing a full chunk for
an empty mask is harmless because `write_pos` does not advance, so the next
store overwrites the garbage.
x86's `MIN_DENSITY_WIDE = 0.05` was set without AVX-512 hardware and with only
two density points measured, so it never probed its own crossover. On Zen 5 it
regressed 4-byte and 8-byte filtering by up to 32%, and AVX2 8-byte by 2.3x.
Both tiers now carry per-(kernel, width) floors, since what sets the floor is
how many lanes one compress covers.
Bands are measured per architecture, interleaved against the same binary with
the kernels disabled, and sit just above the crossover so a core with a
different shuffle-to-scalar balance cannot regress. Verified no regression at
any density on both architectures. Within band:
NEON (Apple M-series): 1 byte 1.4-10.1x, 2 bytes 1.3-9.2x,
4 bytes 1.2-5.1x, 8 bytes 1.3x over 0.50..0.80
x86 (Zen 5, AVX-512): 1 byte 1.9-5.7x, 2 bytes 1.1-3.9x,
4 bytes 1.1-2.8x, 8 bytes 1.1-3.1x
x86 (Zen 5, AVX2 only): 4 bytes 1.1-2.5x, 8 bytes 1.4-2.0x
The 8-byte NEON kernel is the only one needing an upper bound: two lanes per
shuffle never amortizes far, and past ~0.8 the byte-LUT's all-set bulk copies
overtake it.
Also promotes simd_compress to a directory module, and installs mimalloc as
the global allocator in the benchmark. Without it the system allocator's
behaviour at these output sizes dominates and destabilizes divan's
per-iteration timing, swinging identical code by up to 20x.
Checks: cargo +nightly fmt --all --check; cargo clippy -p vortex-array
--all-targets --all-features; cargo nextest run -p vortex-array -p
vortex-buffer -p vortex-mask (4058 passed); cargo test --doc; rustdoc with
-D warnings. Full vortex-array suite also run on the AVX-512 machine (3180
passed), which is the first time the AVX-512 and AVX2 kernels have executed.
Signed-off-by: Robert Kruszewski <robert@spiraldb.com>
Gives 1- and 2-byte elements a vector compress on the AVX2 tier, where they previously had none and fell through to the scalar byte-LUT. The justification I committed for that gap was wrong. It claimed AVX2 has "no cross-lane byte shuffle, so 1/2-byte elements stay on the scalar byte-LUT" - but `pshufb` only has to span as many bytes as the chunk holds, and eight lanes of a 1- or 2-byte element need at most a 16-byte table. One 128-bit lane covers that, so the NEON `tbl` construction ports over unchanged: same LUT layout, same branchless chunk loop, only the load/shuffle/store intrinsics differ. `compress_lut` and `compress_tail` move up to the parent module and are now shared by both. Enabling AVX2 rather than SSSE3 gets the VEX-encoded `vpshufb`, avoiding AVX/SSE transition penalties in mixed code. AVX-512F parts without VBMI2 (Skylake-X) fall through to these too. Measured on Zen 5 with AVX-512 detection forced off, interleaved against the same binary with the kernels disabled: 1 byte, floor 0.15: 1.5x at floor, 2.6x at 0.8 2 bytes, floor 0.25: 1.0x at floor, 2.6x at 0.65 The floors differ despite the identical lane count because the two widths compete against different scalar strategies: `byte_compress_density_threshold` sends 1-byte elements to the byte-LUT at every density on x86, and that LUT is weak on a sparse mask, while 2-byte elements fall to the set-bit walk, which holds out to 0.25. A floor of 0.15 for 2-byte regressed 27%. Retuning that threshold would raise the 1-byte floor to match, and is noted in the docs. Also corrects two numbers in the previous commit that came from a single unreplicated run: the AVX2 8-byte `vpermd` kernel is 1.0x at its 0.45 floor and 1.4x at 0.65, not 1.4x and 2.0x. Two further runs agree. Verified no regression at any density on either x86 tier, and that the shared helper move left the NEON kernels' timings unchanged. `avx2_kernels_match_scalar` now covers the pshufb kernels directly, since a VBMI2 machine's dispatcher never selects them. Checks: cargo +nightly fmt --all --check; cargo clippy -p vortex-array --all-targets --all-features; cargo clippy --target x86_64-unknown-linux-gnu -p vortex-array --lib; cargo nextest run -p vortex-array -p vortex-buffer -p vortex-mask (4058 passed); rustdoc with -D warnings. On the AVX-512 machine: full vortex-array suite (3180 passed) and clippy --all-targets. Signed-off-by: Robert Kruszewski <robert@spiraldb.com>
robert3005
force-pushed
the
rk/filterbuffer
branch
from
July 28, 2026 15:50
be5e458 to
b471442
Compare
robert3005
marked this pull request as ready for review
July 29, 2026 00:18
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.
Port logic from Primitive types to Decimal types and adjust the way we choose
filter strategies from simple threshold for indices and slices