slow-J opened a new pull request, #16631:
URL: https://github.com/apache/lucene/pull/16631
I was looking into HNSW indexing performance and saw in an internal Amazon
JFR profiler
`OffHeapScalarQuantizedVectorValues.unpackNibbles` at ~7% on a full index
build.
Attempting to improve performance by vectorizing it.
Routing it through `VectorUtilSupport` as `int4Unpack` so Panama unpacks a
register of bytes per iteration instead of one byte at a time. The output
should be identical to the previous scalar loop.
Benchmarked via JMH:
Added a JMH test and ran
```
java --module-path lucene/benchmark-jmh/build/benchmarks \
--module org.apache.lucene.benchmark.jmh \
'VectorUtilBenchmark.binaryHalfByteUnpack.*' -p size=128,1024,4096
```
hardware:
EC2 Host: m6g.16xlarge
AWS Graviton2 (Neoverse N1), aarch64, no SVE, preferredBitSize=128; FMA
enabled, JDK 25.0.2
Benchmark (size) Mode Cnt Score
Error Units
VectorUtilBenchmark.binaryHalfByteUnpackScalar 128 thrpt 15 13.562 ±
0.026 ops/us
VectorUtilBenchmark.binaryHalfByteUnpackScalar 1024 thrpt 15 2.157 ±
0.003 ops/us
VectorUtilBenchmark.binaryHalfByteUnpackScalar 4096 thrpt 15 0.553 ±
0.001 ops/us
VectorUtilBenchmark.binaryHalfByteUnpackVector 128 thrpt 15 99.676 ±
1.297 ops/us
VectorUtilBenchmark.binaryHalfByteUnpackVector 1024 thrpt 15 28.519 ±
0.085 ops/us
VectorUtilBenchmark.binaryHalfByteUnpackVector 4096 thrpt 15 8.339 ±
0.217 ops/us
7.35x at 128 dims, 13.2x at 1024, 15.1x at 4096.
I think this may be somewhat related to
https://github.com/apache/lucene/issues/15697
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
