fastlanes: streaming compare + between kernels for BitPacked#8015
Draft
joseph-isaacs wants to merge 1 commit into
Draft
fastlanes: streaming compare + between kernels for BitPacked#8015joseph-isaacs wants to merge 1 commit into
joseph-isaacs wants to merge 1 commit into
Conversation
Adds `CompareKernel` and `BetweenKernel` for `BitPacked` that walk the
encoded array one 1024-element FastLanes block at a time through a single
reused scratch buffer, splice any `Patches` into the block in place via a
sorted-index cursor, then fold a `Fn(T) -> bool` predicate over the block
and write the bits directly into the output bit buffer. The materialised
primitive never appears.
The inner predicate-fold matches the canonical `BitBuffer::collect_bool`
shape — pack 64 bools into a `u64` in a tight loop — which rustc
auto-vectorises into the same `pcmpeq` + `psllq` (vector shift to bit
position) + `por` (OR into accumulator) pattern that `arrow-ord::apply_op`
lowers to. Verified via `objdump` on the bench binary (344 monomorphised
`stream_predicate` variants emit those SIMD instructions in the inner loop).
Smallest possible diff: only adds the two kernels and a private helper
shared between them, no benches, no public-API expansion beyond the two
trait impls.
encodings/fastlanes/public-api.lock | 8 +
encodings/fastlanes/src/bitpacking/compute/between.rs | 248 +
encodings/fastlanes/src/bitpacking/compute/compare.rs | 187 +
encodings/fastlanes/src/bitpacking/compute/mod.rs | 3 +
encodings/fastlanes/src/bitpacking/compute/stream_predicate.rs | 211 +
encodings/fastlanes/src/bitpacking/vtable/kernels.rs | 4 +
6 files changed, 661 insertions(+)
Checks:
- cargo nextest run -p vortex-fastlanes (278 passed)
- cargo clippy -p vortex-fastlanes --all-targets --all-features
- cargo +nightly fmt -p vortex-fastlanes --check
- ./scripts/public-api.sh (only adds the two new trait impls)
Signed-off-by: Claude <noreply@anthropic.com>
joseph-isaacs
added
the
action/benchmark
github-actions
Bot
removed
the
action/benchmark
Contributor
Polar Signals Profiling ResultsLatest Run
Powered by Polar Signals Cloud |
Contributor
Benchmarks: PolarSignals ProfilingVortex (geomean): 1.052x ➖ datafusion / vortex-file-compressed (1.052x ➖, 0↑ 3↓)
|
Contributor
File Sizes: PolarSignals ProfilingNo file size changes detected. |
Contributor
Benchmarks: FineWeb NVMeVerdict: No clear signal (low confidence) datafusion / vortex-file-compressed (0.968x ➖, 1↑ 0↓)
datafusion / vortex-compact (0.982x ➖, 0↑ 0↓)
datafusion / parquet (0.975x ➖, 1↑ 0↓)
duckdb / vortex-file-compressed (0.963x ➖, 2↑ 0↓)
duckdb / vortex-compact (1.001x ➖, 0↑ 0↓)
duckdb / parquet (0.971x ➖, 1↑ 0↓)
Full attributed analysis
|
Contributor
File Sizes: FineWeb NVMeNo file size changes detected. |
Contributor
Benchmarks: TPC-H SF=1 on NVMEVerdict: No clear signal (low confidence) datafusion / vortex-file-compressed (1.043x ➖, 0↑ 1↓)
datafusion / vortex-compact (1.038x ➖, 0↑ 1↓)
datafusion / parquet (0.993x ➖, 2↑ 0↓)
datafusion / arrow (1.052x ➖, 0↑ 4↓)
duckdb / vortex-file-compressed (0.980x ➖, 0↑ 0↓)
duckdb / vortex-compact (1.001x ➖, 0↑ 0↓)
duckdb / parquet (0.992x ➖, 2↑ 1↓)
duckdb / duckdb (0.999x ➖, 0↑ 0↓)
Full attributed analysis
|
Contributor
File Sizes: TPC-H SF=1 on NVMEFile Size Changes (195 files changed, -98.4% overall, 0↑ 195↓)
Totals:
|
Merging this PR will improve performance by 32.36%
|
| Mode | Benchmark | BASE |
HEAD |
Efficiency | |
|---|---|---|---|---|---|
| ⚡ | Simulation | chunked_varbinview_canonical_into[(1000, 10)] |
197.9 µs | 162 µs | +22.19% |
| ⚡ | Simulation | chunked_varbinview_into_canonical[(100, 100)] |
358.4 µs | 323.5 µs | +10.78% |
| ⚡ | Simulation | chunked_varbinview_into_canonical[(1000, 10)] |
211.2 µs | 175.8 µs | +20.11% |
| ⚡ | Simulation | chunked_varbinview_opt_canonical_into[(1000, 10)] |
224.8 µs | 188.6 µs | +19.23% |
| ⚡ | Simulation | fast_lt_out_of_range[16, 1024] |
67.8 µs | 31 µs | ×2.2 |
| ⚡ | Simulation | fast_lt_out_of_range[4, 1024] |
87.5 µs | 37.2 µs | ×2.4 |
| ⚡ | Simulation | baseline_lt[4, 65536] |
251.9 µs | 201 µs | +25.32% |
| ⚡ | Simulation | fast_eq_out_of_range[4, 1024] |
67 µs | 30.4 µs | ×2.2 |
| ⚡ | Simulation | fast_lt_out_of_range[4, 65536] |
262 µs | 109.3 µs | ×2.4 |
| ⚡ | Simulation | new_alp_prim_test_between[f32, 2048] |
62.1 µs | 53.1 µs | +16.9% |
| ❌ | Simulation | baseline_lt[4, 1024] |
64.1 µs | 78.8 µs | -18.64% |
| ⚡ | Simulation | fast_eq_out_of_range[16, 1024] |
67.7 µs | 31.1 µs | ×2.2 |
| ⚡ | Simulation | fast_eq_out_of_range[4, 65536] |
246 µs | 86.9 µs | ×2.8 |
| ⚡ | Simulation | fast_eq_out_of_range[16, 65536] |
291.1 µs | 137.4 µs | ×2.1 |
| ⚡ | Simulation | fast_lt_out_of_range[16, 65536] |
306.3 µs | 126.3 µs | ×2.4 |
| ⚡ | Simulation | baseline_eq[16, 65536] |
259.4 µs | 229.9 µs | +12.8% |
| ⚡ | Simulation | baseline_eq[4, 65536] |
237.9 µs | 185.1 µs | +28.48% |
| ⚡ | Simulation | baseline_lt[16, 65536] |
274.5 µs | 217.7 µs | +26.07% |
| ❌ | Simulation | new_alp_prim_test_between[f32, 32768] |
153.2 µs | 200.9 µs | -23.72% |
| ⚡ | Simulation | new_alp_prim_test_between[f64, 32768] |
250.4 µs | 208.5 µs | +20.1% |
| ... | ... | ... | ... | ... | ... |
ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.
Tip
Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.
Comparing claude/fastlane-compare-kernel-7slGu (ee44dd6) with develop (7b47788)
Contributor
Benchmarks: TPC-DS SF=1 on NVMEVerdict: No clear signal (low confidence) datafusion / vortex-file-compressed (1.052x ➖, 2↑ 9↓)
datafusion / vortex-compact (1.059x ➖, 0↑ 9↓)
datafusion / parquet (1.051x ➖, 0↑ 12↓)
duckdb / vortex-file-compressed (1.051x ➖, 1↑ 21↓)
duckdb / vortex-compact (1.048x ➖, 0↑ 11↓)
duckdb / parquet (1.039x ➖, 0↑ 8↓)
duckdb / duckdb (1.056x ➖, 0↑ 21↓)
Full attributed analysis
|
Contributor
File Sizes: TPC-DS SF=1 on NVMENo file size changes detected. |
Contributor
Benchmarks: FineWeb S3Verdict: No clear signal (environment too noisy confidence) datafusion / vortex-file-compressed (1.049x ➖, 0↑ 1↓)
datafusion / vortex-compact (0.821x ➖, 1↑ 0↓)
datafusion / parquet (0.976x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed (0.996x ➖, 0↑ 0↓)
duckdb / vortex-compact (1.073x ➖, 0↑ 0↓)
duckdb / parquet (0.990x ➖, 0↑ 0↓)
Full attributed analysis
|
Contributor
Benchmarks: Statistical and Population GeneticsVerdict: No clear signal (low confidence) duckdb / vortex-file-compressed (0.992x ➖, 0↑ 0↓)
duckdb / vortex-compact (1.005x ➖, 0↑ 0↓)
duckdb / parquet (1.011x ➖, 0↑ 0↓)
Full attributed analysis
|
Contributor
File Sizes: Statistical and Population GeneticsNo file size changes detected. |
Contributor
Benchmarks: TPC-H SF=10 on NVMEVerdict: No clear signal (low confidence) datafusion / vortex-file-compressed (0.980x ➖, 0↑ 0↓)
datafusion / vortex-compact (0.987x ➖, 0↑ 0↓)
datafusion / parquet (0.985x ➖, 0↑ 0↓)
datafusion / arrow (0.966x ➖, 1↑ 0↓)
duckdb / vortex-file-compressed (0.983x ➖, 0↑ 0↓)
duckdb / vortex-compact (0.994x ➖, 0↑ 0↓)
duckdb / parquet (0.990x ➖, 0↑ 0↓)
duckdb / duckdb (0.997x ➖, 0↑ 0↓)
Full attributed analysis
|
Contributor
File Sizes: TPC-H SF=10 on NVMENo file size changes detected. |
Contributor
Benchmarks: Clickbench on NVMEVerdict: No clear signal (low confidence) datafusion / vortex-file-compressed (0.944x ➖, 4↑ 0↓)
datafusion / parquet (0.941x ➖, 3↑ 0↓)
duckdb / vortex-file-compressed (0.931x ➖, 8↑ 0↓)
duckdb / parquet (0.956x ➖, 4↑ 0↓)
duckdb / duckdb (0.956x ➖, 4↑ 0↓)
Full attributed analysis
|
Contributor
File Sizes: Clickbench on NVMEFile Size Changes (1 files changed, -0.0% overall, 0↑ 1↓)
Totals:
|
Contributor
Benchmarks: Random AccessVortex (geomean): 0.850x ✅ unknown / unknown (0.879x ✅, 21↑ 0↓)
|
Contributor
Benchmarks: TPC-H SF=1 on S3Verdict: No clear signal (environment too noisy confidence) datafusion / vortex-file-compressed (1.060x ➖, 0↑ 0↓)
datafusion / vortex-compact (1.052x ➖, 0↑ 0↓)
datafusion / parquet (1.013x ➖, 0↑ 1↓)
duckdb / vortex-file-compressed (0.996x ➖, 0↑ 0↓)
duckdb / vortex-compact (0.976x ➖, 0↑ 0↓)
duckdb / parquet (1.001x ➖, 0↑ 0↓)
Full attributed analysis
|
Contributor
Benchmarks: CompressionVortex (geomean): 0.997x ➖ unknown / unknown (0.997x ➖, 2↑ 3↓)
|
Contributor
Benchmarks: TPC-H SF=10 on S3Verdict: No clear signal (environment too noisy confidence) datafusion / vortex-file-compressed (0.923x ➖, 2↑ 0↓)
datafusion / vortex-compact (0.893x ➖, 2↑ 0↓)
datafusion / parquet (0.975x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed (0.906x ➖, 0↑ 0↓)
duckdb / vortex-compact (0.956x ➖, 0↑ 0↓)
duckdb / parquet (0.867x ➖, 0↑ 0↓)
Full attributed analysis
|
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
2 participants
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.
Adds
CompareKernelandBetweenKernelforBitPackedthat walk theencoded array one 1024-element FastLanes block at a time through a single
reused scratch buffer, splice any
Patchesinto the block in place via asorted-index cursor, then fold a
Fn(T) -> boolpredicate over the blockand write the bits directly into the output bit buffer. The materialised
primitive never appears.
The inner predicate-fold matches the canonical
BitBuffer::collect_boolshape — pack 64 bools into a
u64in a tight loop — which rustcauto-vectorises into the same
pcmpeq+psllq(vector shift to bitposition) +
por(OR into accumulator) pattern thatarrow-ord::apply_oplowers to. Verified via
objdumpon the bench binary (344 monomorphisedstream_predicatevariants emit those SIMD instructions in the inner loop).Smallest possible diff: only adds the two kernels and a private helper
shared between them, no benches, no public-API expansion beyond the two
trait impls.
Checks:
Signed-off-by: Claude noreply@anthropic.com