Remove the complicated block search logic for a simpler branchless #1124
Conversation
Codecov Report
@@ Coverage Diff @@
## main #1124 +/- ##
=======================================
Coverage 89.50% 89.51%
=======================================
Files 203 203
Lines 20231 20201 -30
=======================================
- Hits 18107 18082 -25
+ Misses 2124 2119 -5
Continue to review full report at Codecov.
|
PSeitz
left a comment
PSeitz
left a comment
There was a problem hiding this comment.
Simpler and faster, really cool. One minor comment. Also I don't know if the data is monotonically increasing or if we can have duplicates.
src/postings/block_search.rs
Outdated
| 1041, 1048, 1061, 1062, 1069, 1072, 1079, 1082, 1083, 1102, 1140, 1165, 1170, 1176, | ||
| 1189, 1195, 1198, 1200, 1201, 1208, 1223, 1240, 1252, 1276, 1307, | ||
| ]; | ||
| for target in block.iter().flat_map(|&el| vec![el - 1, el].into_iter()) { |
There was a problem hiding this comment.
I would add
vec![el - 1, el, el+1] to test with a jump size larger than 1 to the next bucket (except the last element .take(block.len() * 3 -1) )
There was a problem hiding this comment.
el +1 is the same as el-1 for the precedent element.
There was a problem hiding this comment.
I'll test all elements in [0..1308).
There was a problem hiding this comment.
Simpler and faster, really cool. One minor comment. Also I don't know if the data is monotonically increasing or if we can have duplicates.
Good point I modified the contract. There was an error in it too.
binary search The code is simpler and faster. Before test postings::bench::bench_segment_intersection ... bench: 2,093,697 ns/iter (+/- 115,509) test postings::bench::bench_skip_next_p01 ... bench: 58,585 ns/iter (+/- 796) test postings::bench::bench_skip_next_p1 ... bench: 160,872 ns/iter (+/- 5,164) test postings::bench::bench_skip_next_p10 ... bench: 615,229 ns/iter (+/- 25,108) test postings::bench::bench_skip_next_p90 ... bench: 1,120,509 ns/iter (+/- 22,271) After test postings::bench::bench_segment_intersection ... bench: 1,747,726 ns/iter (+/- 52,867) test postings::bench::bench_skip_next_p01 ... bench: 55,205 ns/iter (+/- 714) test postings::bench::bench_skip_next_p1 ... bench: 131,433 ns/iter (+/- 2,814) test postings::bench::bench_skip_next_p10 ... bench: 478,830 ns/iter (+/- 12,794) test postings::bench::bench_skip_next_p90 ... bench: 931,082 ns/iter (+/- 31,468)
binary search
The code is simpler and faster.