-
Notifications
You must be signed in to change notification settings - Fork 4k
ARROW-12032: [Rust] Optimize comparison kernels #9759
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Closed
Closed
Changes from 5 commits
Commits
Show all changes
7 commits
Select commit
Hold shift + click to select a range
f233f47
Use trusted_len for comparison kernels
Dandandan 1aa7baf
Fmt
Dandandan 5f2dfa2
Add documentation and doctest
Dandandan c4cb72a
Add documentation
Dandandan 0e1c8ed
Clippy
Dandandan d51e75e
Fix benches
Dandandan aa4a4fb
Add faster primitive versions for comparison kernels
Dandandan File filter
Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
There are no files selected for viewing
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
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
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Outdated suggestions cannot be applied.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.
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I wonder if the bool iterator could be split into chunks (for example, using https://docs.rs/itertools/0.4.2/itertools/struct.Chunks.html or alternatively using https://doc.rust-lang.org/std/primitive.slice.html#method.chunks) of 8 bool values, then each chunk is mapped into a byte by converting each bool value into a byte (for example using std::mem::transmute::<bool, u8>), then shifting according to the position in the chunk, and applying in the output byte, and finally the resulting byte iterator would be used to build the buffer directly. This is the fastest implementation I can imagine because it eliminates as many conditions / checks as possible (and conditions are the enemy of fast).
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I also think there are some faster ways to speed up the inner loop, yours sounds like a great idea to try out. I was also looking at the arrow2 repository of @jorgecarleitao , but I think I have been looking to an older commit before which turned out to be slower (I expected it to be faster, but sometimes the compiler can be quite surprising in what compiles to efficient code.
I think the latest version is over here:
https://github.com/jorgecarleitao/arrow2/blob/be905f1b1f0293ef427387bc35b2e9956ec3336f/src/bitmap/mutable.rs#L209
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I agree with you all ❤️
I admit I have spent an immoral amount of time trying to optimize bitmaps, but I have unfortunately not yet concluded what is the best way to handle them. I think that we may not being able to express to the compiler what we want it to do (some kind of operation over a single byte). @yordan-pavlov suggestion is a great one in that direction, though.
FWIW, on my computer (a VM on azure), arrow master (not this PR) is giving
and
arrow2is givingThis PR's idea on arrow2 (with corresponding changes) is giving me
-14%oneq Float32and+35%oneq scalar Float32. I pushed these benches to master there.Note the difference between scalar and non-scalar: it is the exact same code on the
trusted_lenfunction, but a 30% difference in performance between them; imo this indicates that we are fighting with the compiler to try to explain what we are trying to express here.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
FYI, I wrote a minimal repo to evaluate these things and added a reddit post to try to get some help / feedback on this.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@yordan-pavlov's idea yields
-50%on array-to-scalar and-10%on array-to-array 🚀There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@yordan-pavlov
I am not sure where the win would be in that case.
I would expect the first idea to be compiled to roughly the same code (if all compiler optimizations work out)?
For the
Vecone - I would expect that would be slower as it introduces an extra loop / allocation and barrier for optimization?There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@Dandandan yes, I agree that it's counter-intuitive, but I find that the compiler often surprises me so it's best to try every option; I will try to extend the benchmark repo when I have a bit more free time later today
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
You are right, sometimes the result can be surprising.
I tried this variation, this compiles to the same unrolled 38 instructions:
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
yes, I tried to benchmark doing the comparison separately, but it's not faster; on my machine the fastest version is:
and that's even faster than:
this is with the test data configured as:
I think 2000 items (the old length of the test data) is much too small for realistic benchmarking, and it would make more sense to benchmark with test data with length same as the default batch size in DataFusion (I think this was recently increased).
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@yordan-pavlov
Also see findings in the PR here (with updated benchmark).
(x * x + x) % 2will give 0 on even and 1 on uneven inputs I, so the pattern/branches will be very predictable, especially in the scalar version.jorgecarleitao/arrow2#17