benwtrent merged PR #13321:
URL: https://github.com/apache/lucene/pull/13321
--
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.a
uschindler commented on PR #13321:
URL: https://github.com/apache/lucene/pull/13321#issuecomment-2088443326
To me changes look fine.
For discussion: In my opinion the conditional code falling back to defaults
should possibly be moved to the VectorUtil class and i think them the boolea
benwtrent commented on PR #13321:
URL: https://github.com/apache/lucene/pull/13321#issuecomment-2088378164
@uschindler I addressed your comments.
--
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 t
ChrisHegarty commented on code in PR #13321:
URL: https://github.com/apache/lucene/pull/13321#discussion_r1582893750
##
lucene/core/src/java21/org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport.java:
##
@@ -390,22 +392,202 @@ private int dotProductBody128(byte[] a,
benwtrent commented on code in PR #13321:
URL: https://github.com/apache/lucene/pull/13321#discussion_r1581482527
##
lucene/core/src/java21/org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport.java:
##
@@ -390,22 +391,139 @@ private int dotProductBody128(byte[] a, by
uschindler commented on code in PR #13321:
URL: https://github.com/apache/lucene/pull/13321#discussion_r1581305081
##
lucene/core/src/java21/org/apache/lucene/internal/vectorization/PanamaVectorUtilSupport.java:
##
@@ -390,22 +391,139 @@ private int dotProductBody128(byte[] a, b
benwtrent commented on PR #13321:
URL: https://github.com/apache/lucene/pull/13321#issuecomment-2079695750
OK, I ran on Google's ARM machine (`Tau T2A machine series`) to make sure
the ARM performance improvements still exist for int4 (and it wasn't some silly
macos thing):
```
Be
ChrisHegarty commented on PR #13321:
URL: https://github.com/apache/lucene/pull/13321#issuecomment-2079213000
I went over the Panama code again, and confused myself about the potential
to overflow - there is no issue, but it was not obvious as one has to match the
number of potential values
rmuir commented on PR #13321:
URL: https://github.com/apache/lucene/pull/13321#issuecomment-2078163094
panama logic looks good to me. Not for this PR, just a question, is it
possible to use this same trick for 8 bit integer calculations too? I imagine
this trick could probably help there as
ChrisHegarty commented on PR #13321:
URL: https://github.com/apache/lucene/pull/13321#issuecomment-2077675140
> Because this makes things measurably slower. ..
Ok. That's a very good reason! ;-)
> And I am not sure how it makes the comparisons easier. We then have to add
LoC to
benwtrent commented on code in PR #13321:
URL: https://github.com/apache/lucene/pull/13321#discussion_r1579422728
##
lucene/core/src/java/org/apache/lucene/util/VectorUtil.java:
##
@@ -179,7 +179,30 @@ public static int int4DotProduct(byte[] a, byte[] b) {
if (a.length != b
ChrisHegarty commented on code in PR #13321:
URL: https://github.com/apache/lucene/pull/13321#discussion_r1579117412
##
lucene/core/src/java/org/apache/lucene/util/VectorUtil.java:
##
@@ -179,7 +179,30 @@ public static int int4DotProduct(byte[] a, byte[] b) {
if (a.length !
benwtrent commented on code in PR #13321:
URL: https://github.com/apache/lucene/pull/13321#discussion_r1577962121
##
lucene/core/src/test/org/apache/lucene/codecs/lucene99/TestLucene99ScalarQuantizedVectorScorer.java:
##
@@ -0,0 +1,145 @@
+/*
+ * Licensed to the Apache Software
ChrisHegarty commented on code in PR #13321:
URL: https://github.com/apache/lucene/pull/13321#discussion_r1577918752
##
lucene/core/src/java/org/apache/lucene/util/hnsw/RandomAccessVectorValues.java:
##
@@ -72,6 +84,11 @@ interface Floats extends RandomAccessVectorValues {
benwtrent commented on code in PR #13321:
URL: https://github.com/apache/lucene/pull/13321#discussion_r1576351827
##
lucene/core/src/java/org/apache/lucene/codecs/lucene99/Lucene99ScalarQuantizedVectorScorer.java:
##
@@ -0,0 +1,283 @@
+/*
+ * Licensed to the Apache Software Foun
benwtrent opened a new pull request, #13321:
URL: https://github.com/apache/lucene/pull/13321
This updates the int4 dot-product comparison to have an optimized one for
when one of the vectors are compressed (the most common search case). This
change actually makes the compressed search on A
16 matches
Mail list logo
