faster iandnot between bitmap containers and run containers #418
Conversation
| wordRangeStart := rc.iv[0].start / 64 | ||
| wordRangeEnd := (rc.iv[len(rc.iv)-1].last()) / 64 // inclusive | ||
|
|
||
| cardinalityChange := popcntSlice(bc.bitmap[wordRangeStart : wordRangeEnd+1]) // before cardinality - after cardinality (for word range) |
There was a problem hiding this comment.
I do not understand this part. I don't understand how you can update the cardinality with a single popcntSlice.
I suspect it is not correct... I am troubled by the fact that our test pass... but my suspicions is that the code is not correct. Can you convince me?
There was a problem hiding this comment.
Oh. Wait. I see what you are doing. Ok. It looks good.
|
This is good and I expect to merge. I will give it some time in case someone wants to review it further. |
|
@lemire I just copied this change over to a benchmarking tool I have that tests a prod workload that uses this function and it seems to break something -- could you hold on merging until I get to the bottom what's happening? |
|
draft mode |
neena
force-pushed
the
neena/faster-iandnot-bm
branch
from
1c6a47d
to
ce0ef6c
Compare
|
all fixed! thanks for your patience |
|
Merging. |
Rather than converting a run container to a bitmap container and then performing iandnot, call removeRange directly.