Mirror/MarginaliaSearch
1
0
Fork 0
mirror of https://github.com/MarginaliaSearch/MarginaliaSearch.git synced 2025-02-24 05:18:58 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

(array, experimental) Add benchmark results for quicksort

Browse Source
This commit is contained in:
Viktor Lofgren 2024-05-14 18:15:30 +02:00
parent 3549be216f
commit f48cf77c4d
1 changed files with 22 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
code/libraries/array/cpp/readme.md
View File

@ -1,12 +1,29 @@
# Array C++ Helpers # LongArray C++ Helpers
This package contains helper functions for working with LongArray objects, This package contains helper functions for working with LongArray objects,
as native C++ calls. The helpers are only built on Linux, and if they are absent, as native C++ calls. The helpers are only built on Linux, and if they are absent,
Java substitutes should be used instead. Java substitutes should be used instead.
Note that the C++ helpers are compiled with march=native, so they are not portable Library loading and access is available through the
[NativeAlgos](java/nu/marginalia/NativeAlgos.java) class.
Note that the C++ helpers are compiled with march=native by default, so they are not portable
between different CPU architectures. between different CPU architectures.
This provides a speedup for some operations, especially when using MemorySegment's This provides a speedup for some operations, especially when compared to using MemorySegment's
get methods and not Unsafe. Library loading and access is available through the get methods and not Unsafe.
[NativeAlgos](java/nu/marginalia/NativeAlgos.java) class.
## Benchmarks
JMH benchmark results for 64 bit quicksort N=2^10, 25 trials, graalvm 21.0.3:
| Implementation | Throughput ops/s | Error p=99% |
|----------------|------------------|-------------|
| native | 114.877 | 4.216 |
| unsafe | 98.443 | 2.723 |
| memorysegment | 72.492 | 1.980 |
Note that odds are stacked in favor of the native implementation
since the dataset is fairly large, meaning the call overhead is
less significant. For smaller datasets, the overhead of native
calls may be a bigger factor.