[MOD-7369] Add AVX512VL flag to FP16 flag to enhance performance. #522
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
With this flag the dimension limit to chose the advanced opt is not required. Expanded low dim bm spaces range.
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #522 +/- ##
=======================================
Coverage 97.13% 97.13%
=======================================
Files 94 94
Lines 4886 4886
=======================================
Hits 4746 4746
Misses 140 140 ☔ View full report in Codecov by Sentry. |
2 tasks
GuyAv46
reviewed
Aug 13, 2024
There was a problem hiding this comment.
How is this file and its functions completely new? Shouldn’t we have other functions to rename or remove?
There was a problem hiding this comment.
I think i fixed it
meiravgri
changed the title
meiravgri
changed the title
alonre24
approved these changes
Aug 13, 2024
github-actions bot
pushed a commit
that referenced
this pull request
Aug 13, 2024
* add AVX512VL flag to FP16 flag to enhance performance. With this flag the dimension limit to chose the advanced opt is not required. Expanded low dim bm spaces range. * add VL to building with prompt in compilation * remove old file (cherry picked from commit 3d77d23)
|
Successfully created backport PR for |
meiravgri
added a commit
that referenced
this pull request
Aug 13, 2024
…e. (#524) [MOD-7369] Add AVX512VL flag to FP16 flag to enhance performance. (#522) * add AVX512VL flag to FP16 flag to enhance performance. With this flag the dimension limit to chose the advanced opt is not required. Expanded low dim bm spaces range. * add VL to building with prompt in compilation * remove old file (cherry picked from commit 3d77d23) Co-authored-by: meiravgri <[email protected]>
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.
Background
In PR #477 we introduced distance functions leveraging
AVX512FP16intrinsic to enhanceFLOAT16distance calculations.We observed that
_mm512_reduce_add_phincurs significant overhead, leading to a substantial degradation in distance calculation performance compared to alternative implementations. However, this overhead becomes less significant at higher dimensions. Therefore,AVX512FP16instruction set was enabled only for dimensions exceeding a certain threshold.Current PR
This PR adds the
-mavx512vlcompilation flag to mitigate the latency issues observed with_mm512_reduce_add_ph.Explanation:
The reduce instruction is a complex instruction, translating into a sequence of instructions under the hood.
With the flag, the compiler uses instructions: 128/256-bit
vaddph. Without the flag, the compiler uses instructions:vaddsh.As demonstrated in the attached performance graph, the first option leads to improved performance across all dimensions.
Additional changes
Expand LOW_DIM spaces benchmarks range. Starts from dim = 55 instead of 100. Still satisfies a no-residual dimension for all types (at dim = 160)