neilconway opened a new pull request, #11992:
URL: https://github.com/apache/arrow-rs/pull/11992
# Which issue does this PR close?
- Split out from #11656
- Related to #11770
# Rationale for this change
This PR applies the optimization from #11770 to an additional code path:
when `MutableArrayData` copies string views into an array that already has data
buffers, the buffer index in the copied views must be updated. We previously
did that by unpacking the raw view (`u128` -> `ByteView`), adding the buffer
index offset, and repacking the view. It is faster to do this by using bitwise
shifts and addition directly. This also avoids a data-dependent branch, which
seems to have been poorly predicted on x86.
Benchmarks:
x86 (AMD EPYC Milan):
- zip_kernels, string views, 8192 rows of 3 to 19 bytes
- array vs scalar
- all_true: 31.19 µs → 6.41 µs (−79.4%)
- 99pct_true: 35.84 µs → 10.94 µs (−69.5%)
- 90pct_true: 63.87 µs → 39.54 µs (−38.1%)
- 50pct_true: 129.4 µs → 113.1 µs (−12.6%)
- true_then_false: 55.02 µs → 43.06 µs (−21.7%)
- 10pct_true: 96.03 µs → 92.53 µs (−3.6%)
- 1pct_true: 81.66 µs → 82.20 µs (+0.7%)
- all_false: 78.78 µs → 81.76 µs (+3.8%) (runs 6.7% apart)
- 50pct_nulls: 112.6 µs → 107.5 µs (−4.5%) (runs 6.4% apart)
- scalar vs array
- all_true: 82.50 µs → 85.26 µs (+3.3%) (runs 4.4% apart)
- 99pct_true: 86.52 µs → 87.81 µs (+1.5%) (runs 4.4% apart)
- 90pct_true: 100.9 µs → 112 µs (+11.0%) (runs 31.5% apart)
- 50pct_true: 136.6 µs → 121 µs (−11.4%) (runs 5.6% apart)
- true_then_false: 58.65 µs → 47.93 µs (−18.3%) (runs 7.2% apart)
- 10pct_true: 69.18 µs → 44.68 µs (−35.4%) (runs 5.1% apart)
- 1pct_true: 39.66 µs → 15.05 µs (−62.1%) (runs 18.5% apart)
- all_false: 30.34 µs → 7.15 µs (−76.4%) (runs 17.9% apart)
- 50pct_nulls: 102.3 µs → 81.27 µs (−20.6%) (runs 8.6% apart)
- array vs array
- all_true: 30.94 µs → 6.82 µs (−77.9%) (runs 2.6% apart)
- 99pct_true: 37.47 µs → 12.13 µs (−67.6%) (runs 2.8% apart)
- 90pct_true: 71.25 µs → 45.33 µs (−36.4%) (runs 13.0% apart)
- 50pct_true: 77.46 µs → 79.37 µs (+2.5%) (runs 2.9% apart)
- true_then_false: 31.60 µs → 7.00 µs (−77.9%) (runs 3.5% apart)
- 10pct_true: 72.32 µs → 47.72 µs (−34.0%) (runs 25.3% apart)
- 1pct_true: 37.44 µs → 12.04 µs (−67.8%) (runs 5.9% apart)
- all_false: 31.26 µs → 7.01 µs (−77.6%) (runs 4.9% apart)
- 50pct_nulls: 77.66 µs → 78.57 µs (+1.2%)
- mutable_array
- mutable str_view 1024: 294.5 µs → 267.7 µs (−9.1%)
- mutable str_view short slices 1024: 19.84 µs → 17.69 µs (−10.9%) (runs
4.3% apart)
ARM (M4 Max):
- zip_kernels, string views, 8192 rows of 3 to 19 bytes
- array vs scalar
- all_true: 4.27 µs → 3.55 µs (−16.9%) (runs 2.9% apart)
- 99pct_true: 7.22 µs → 6.61 µs (−8.4%) (runs 5.6% apart)
- 90pct_true: 23.69 µs → 22.70 µs (−4.2%)
- 50pct_true: 63.98 µs → 64.23 µs (+0.4%)
- true_then_false: 25.21 µs → 25.53 µs (+1.3%) (runs 3.9% apart)
- 10pct_true: 53.92 µs → 56.04 µs (+3.9%) (runs 6.1% apart)
- 1pct_true: 49.47 µs → 51.00 µs (+3.1%) (runs 4.0% apart)
- all_false: 45.95 µs → 46.10 µs (+0.3%)
- 50pct_nulls: 60.34 µs → 60.33 µs (−0.0%)
- scalar vs array
- all_true: 49.63 µs → 50.10 µs (+0.9%)
- 99pct_true: 53.14 µs → 53.63 µs (+0.9%)
- 90pct_true: 57.90 µs → 58.62 µs (+1.2%)
- 50pct_true: 68.40 µs → 68.32 µs (−0.1%)
- true_then_false: 28.06 µs → 28.33 µs (+1.0%) (runs 3.7% apart)
- 10pct_true: 27.70 µs → 27.16 µs (−2.0%) (runs 3.4% apart)
- 1pct_true: 10.63 µs → 10.11 µs (−4.9%) (runs 6.1% apart)
- all_false: 4.22 µs → 3.58 µs (−15.3%) (runs 6.1% apart)
- 50pct_nulls: 47.07 µs → 46.42 µs (−1.4%) (runs 2.7% apart)
- array vs array
- all_true: 4.40 µs → 3.59 µs (−18.3%)
- 99pct_true: 7.51 µs → 6.97 µs (−7.2%) (runs 7.4% apart)
- 90pct_true: 24.69 µs → 24.15 µs (−2.2%) (runs 7.6% apart)
- 50pct_true: 49.95 µs → 51.71 µs (+3.5%) (runs 11.4% apart)
- true_then_false: 4.50 µs → 3.64 µs (−18.9%)
- 10pct_true: 24.68 µs → 23.40 µs (−5.2%)
- 1pct_true: 7.63 µs → 6.79 µs (−11.1%) (runs 3.1% apart)
- all_false: 4.38 µs → 3.60 µs (−17.7%) (runs 2.2% apart)
- 50pct_nulls: 49.16 µs → 50.45 µs (+2.6%) (runs 4.2% apart)
- mutable_array
- mutable str_view 1024: 186.4 µs → 156.2 µs (−16.2%) (runs 6.3% apart)
- mutable str_view short slices 1024: 10.49 µs → 10.04 µs (−4.3%) (runs
4.7% apart)
# What changes are included in this PR?
See above.
# Are these changes tested?
Yes; existing tests pass, new test added.
# Are there any significant user-facing changes?
No.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
