neoremind commented on PR #16631:
URL: https://github.com/apache/lucene/pull/16631#issuecomment-5566047817

   Nice optimization! In term of the e2e benchmark, a minor suggestion: maybe 
include a profile showing what portion this targets? enabling JFR to capture 
CPU time spending, and I usually use this command to convert it to a flamegraph 
for visualization. 
   
   ```
   async-profiler-4.x/bin/jfrconv --cpu-time -o html knn-perf-test.jfr 
flamegraph.html
   ```
   
   I have a quick run on the baseline, the unpack showed as single-digit % CPU 
time as you came up with, after this optimization, maybe it should disappear 
completely.
   
   In addition, I noticed the dominating cost is the int4 dot product, my 
question (maybe my thought is naive): we materialize to unpacked byte[] and the 
in the dot-product call path, we load them into vectorized form factor again, 
is that intermediate transformation necessary? is it possible to do in 
bothPacked way?
   
   


-- 
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]

Reply via email to