Implement off-heap scoring for OSQ 4, 7, and 8 bit representations #15257
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@mccullocht maybe we only do the byte part of the comparisons off-heap? Then apply the corrections all on heap. I would assume applying corrections is pretty cheap, but even then if we did it in bulk, maybe on-heap bulk correction application is pretty fast? |
|
I did move just the vector dot product off-heap and I'm not planning to do anything clever with the corrections. I'm not sure that would pay off anyway -- you'd have to transpose from row view to column view to parallelize that work, and it would be 128-bit on x86 which may not go well. I was assuming that accessing the corrective terms was messing with performance but larger jfr stacks point at a more mysterious culprit. This PR spends more time in lane reduction (???) and 128-bit loads of data from the memory segment (probably memory latency). For the latter case its weird that I don't see anything on the baseline when I know that copying to the heap should be inducing a similar hit. |
|
@mccullocht you might find this interesting: #15272 |
|
I plan to try the simplest thing first and just copy the dot product code for byte[] x MemorySegment to see if that yields an improvement, then go from there. I hoped that the JVM would monomorphize these calls but I guess not. |
mccullocht
force-pushed
the
osq-offheap-vector
branch
from
9e1d76e to
e7178bc
Compare
|
Ok, repeated the experiment with a specific byte[] x MemorySegment implementation. In the luceneutil benchmarks I'm not suffering from the same inlining/pollution issues. The profiles remain the same where suddenly |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Partial implementation of #15155
So far this is not any faster than the alternative. On an AMD RYZEN AI MAX+ 395
DO NOT MERGE
Performance observations: on an avx512 host the profiles are quite different. the original path spends a most time in dotProductBody512 followed by Int512Vector.reduceLanes(). the new path spends much more time in reduceLanes() but also spends more time loading from a memory segment for the input vectors -- a 128 bit load from a memory segment instead of a heap array. this could be memory latency but in that case why doesn't the load into the heap array show up in the profile?