jpountz commented on PR #13658:
URL: https://github.com/apache/lucene/pull/13658#issuecomment-2302929875
I ran wikibigall on a M3, which is interesting because it does inline the
splitLongs call, both before and after the change (presumably because the
generated native code is smaller thus still candidate for inlining), and the
benchmark doesn't get a slowdown (but neither a speedup or a very small one):
```
TaskQPS baseline StdDevQPS
my_modified_version StdDev Pct diff p-value
OrHighLow 1387.75 (4.7%) 1358.50
(6.4%) -2.1% ( -12% - 9%) 0.236
AndStopWords 90.68 (5.1%) 88.85
(6.6%) -2.0% ( -13% - 10%) 0.278
And3Terms 438.98 (4.6%) 433.21
(5.7%) -1.3% ( -11% - 9%) 0.425
HighTermTitleSort 167.18 (4.6%) 165.66
(4.3%) -0.9% ( -9% - 8%) 0.519
HighTermTitleBDVSort 82.36 (4.3%) 81.75
(5.1%) -0.7% ( -9% - 9%) 0.618
OrHighNotHigh 624.72 (4.5%) 620.27
(5.1%) -0.7% ( -9% - 9%) 0.642
AndHighHigh 222.77 (3.9%) 221.34
(4.1%) -0.6% ( -8% - 7%) 0.611
Or2Terms2StopWords 424.69 (4.7%) 422.05
(5.3%) -0.6% ( -10% - 9%) 0.695
AndHighMed 426.20 (4.9%) 423.75
(5.0%) -0.6% ( -10% - 9%) 0.715
Or3Terms 424.04 (4.8%) 421.69
(5.0%) -0.6% ( -9% - 9%) 0.721
OrHighRare 486.01 (5.4%) 483.99
(5.8%) -0.4% ( -10% - 11%) 0.813
OrHighMed 513.21 (4.9%) 511.77
(4.8%) -0.3% ( -9% - 9%) 0.854
PKLookup 409.03 (6.5%) 408.23
(7.2%) -0.2% ( -12% - 14%) 0.928
HighTermMonthSort 1781.87 (4.0%) 1780.09
(6.2%) -0.1% ( -9% - 10%) 0.952
OrNotHighHigh 570.80 (5.0%) 570.42
(7.1%) -0.1% ( -11% - 12%) 0.973
HighTermDayOfYearSort 1509.01 (5.9%) 1512.74
(6.1%) 0.2% ( -11% - 12%) 0.896
OrNotHighMed 686.50 (5.2%) 690.48
(7.7%) 0.6% ( -11% - 14%) 0.781
OrHighNotMed 784.76 (5.2%) 789.88
(4.8%) 0.7% ( -8% - 11%) 0.683
CountAndHighMed 507.69 (4.4%) 511.31
(5.8%) 0.7% ( -9% - 11%) 0.664
OrHighHigh 193.31 (4.8%) 195.18
(4.3%) 1.0% ( -7% - 10%) 0.506
TermDTSort 526.16 (3.1%) 531.24
(5.7%) 1.0% ( -7% - 10%) 0.508
And2Terms2StopWords 449.28 (4.5%) 454.04
(5.5%) 1.1% ( -8% - 11%) 0.506
AndHighLow 2221.82 (6.3%) 2255.93
(7.5%) 1.5% ( -11% - 16%) 0.484
CountAndHighHigh 188.97 (4.4%) 191.99
(4.9%) 1.6% ( -7% - 11%) 0.278
OrHighNotLow 1001.86 (6.4%) 1021.97
(6.2%) 2.0% ( -9% - 15%) 0.314
OrStopWords 100.73 (4.0%) 102.78
(6.4%) 2.0% ( -8% - 12%) 0.229
HighTerm 914.24 (5.6%) 933.01
(6.2%) 2.1% ( -9% - 14%) 0.271
MedTerm 1122.89 (5.2%) 1146.58
(7.2%) 2.1% ( -9% - 15%) 0.287
LowTerm 1617.38 (7.5%) 1661.67
(6.7%) 2.7% ( -10% - 18%) 0.221
OrNotHighLow 2220.24 (5.5%) 2294.40
(8.9%) 3.3% ( -10% - 18%) 0.152
CountOrHighMed 461.03 (5.7%) 477.71
(6.4%) 3.6% ( -7% - 16%) 0.058
CountOrHighHigh 265.96 (6.2%) 279.21
(7.7%) 5.0% ( -8% - 20%) 0.024
```
I'm debating whether I should revert this change or whether we should keep
it because the micro benchmark suggests it's a speedup (both on my Ryzen and my
M3) and we might find ways to take advantage of this speedup later on. I'm
curious if you have an opinion @gsmiller.
--
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: issues-unsubscr...@lucene.apache.org
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org
