jbellis closed pull request #12303: Address HNSW Searcher performance regression
URL: https://github.com/apache/lucene/pull/12303
--
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.
gsmiller commented on issue #12317:
URL: https://github.com/apache/lucene/issues/12317#issuecomment-1556202535
I'm no expert in this area of our codec, but I'm curious to understand the
issue a bit better. From the flame chart you provide, it looks like you're
primarily looking at an indexi
rmuir commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556224675
i made the benchmarks easier to run with something like this:
```
git clone https://github.com/rmuir/vectorbench
cd vectorbench
mvn verify
java -jar target/vectorbench.jar
ChrisHegarty commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556278132
I didn't get an anywhere with Luceneutil yet! :-( (I haven't been able to
run it successfully, getting OOM errors )
--
This is an automated message from the Apache Git Service.
rmuir commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556284488
thanks for sanity checking! i'm still working on the repo and making
improvements. would be super-curious if you could 'git pull' and try
-psize=1024 on your avx512 machine. hopefully it l
rmuir commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556285923
With latest commits to that vectorbench I see this on my m1:
```
Benchmark (size) Mode Cnt Score Error
Units
DotProductBenchmark.dotProductNew
rmuir commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556287294
thanks, glad it fixes the problem. i am running it across all the sizes we
test and seeing how it looks on both my machines.
--
This is an automated message from the Apache Git Service.
ChrisHegarty commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556288076
> we were being inefficient.
If I understand this correctly, the inefficiency was too many reduceLances,
right? You replaced it with addition of the accumulators before reduc
rmuir commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556288827
yes, i think we have to imagine it as a scalar operation that gets slower as
vector size increases. i looked into it and read this answer and changed the
code:
https://stackoverflow.com/q
rmuir commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556297297
i pushed one more commit to improve for "unaligned" vectors. the way to
think about it, with unrolling, we do 64-at-a-time on avx512.
So it isn't good to do worst-case 63 scalar com
rmuir commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556297355
aarch64:
```
Benchmark (size) Mode CntScore Error
Units
DotProductBenchmark.dotProductNew 1 thrpt5 322.255 ± 0.496
ops/us
rmuir commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556297561
skylake:
```
Benchmark (size) Mode CntScore Error
Units
DotProductBenchmark.dotProductNew 1 thrpt5 153.702 ± 2.576
ops/us
rmuir commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556403183
I pushed a new benchmark to https://github.com/rmuir/vectorbench for the
binary dot product.
Basically this has to act like:
```
int sum = 0;
for (...) {
short product
rmuir commented on PR #12311:
URL: https://github.com/apache/lucene/pull/12311#issuecomment-1556405150
and here's the results on my aarch64 mac, which has only 128-bit vectors and
gets that disappointing generic impl:
```
Benchmark(size) Mode CntS
tang-hi opened a new pull request, #12322:
URL: https://github.com/apache/lucene/pull/12322
### Description
solve the bug that @msokolov mentioned in
[PR](https://github.com/apache/lucene/pull/12255#issuecomment-1553088549)
--
This is an automated message from the Apache Git Servic
zhaih merged PR #12257:
URL: https://github.com/apache/lucene/pull/12257
--
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.apach
zhaih commented on code in PR #12246:
URL: https://github.com/apache/lucene/pull/12246#discussion_r1199956128
##
lucene/analysis/common/src/java/org/apache/lucene/analysis/synonym/word2vec/Word2VecSynonymProvider.java:
##
@@ -42,6 +42,7 @@ public class Word2VecSynonymProvider {
17 matches
Mail list logo
