Open
Conversation
PointKernel
added
the
type: improvement
sleeepyjack
approved these changes
Mar 22, 2024
Collaborator
sleeepyjack
left a comment
sleeepyjack
left a comment
There was a problem hiding this comment.
Looks good. Shall I run some benchmarks comparing perf on H100?
Member
Author
Please do. I expect a maximum difference of 0.5% to 1%. |
Collaborator
|
🙅 Something's not right: |
Member
Author
|
About 50% slower for |
sleeepyjack
requested changes
Mar 23, 2024
Collaborator
sleeepyjack
left a comment
sleeepyjack
left a comment
There was a problem hiding this comment.
Not sure if the perf regression is worth investigating. If we can create a minimal repro we can ping our CG devs. I'm also ok with shelving the PR for now.
Member
Author
|
I cannot reproduce the performance regression with my local RTX8000: Could this issue be H100-specific? |
Collaborator
|
That's expected on a <sm_90 arch. From sm_90 going forward the function will use a different code path leveraging the new |
PointKernel
added
Needs Review
Collaborator
|
Update: I ran the same exact benchmarks on another H100 node (CTK 12.3) today and wasn't able to reproduce the regression I initially reported. Will run another test on a different node and report back. |
Member
Author
|
@sleeepyjack Any updates on H100 perf results? |
sleeepyjack
reviewed
May 29, 2024
Comment on lines
81
to
+92
| auto const tile = cg::tiled_partition<CGSize>(block); | ||
| auto const found = ref.find(tile, key); | ||
|
|
||
| #if defined(CUCO_HAS_CG_INVOKE_ONE) | ||
| cg::invoke_one(tile, [&]() { | ||
| *(output_begin + idx) = found == ref.end() ? ref.empty_key_sentinel() : *found; | ||
| }); | ||
| #else | ||
| if (tile.thread_rank() == 0) { | ||
| *(output_begin + idx) = found == ref.end() ? ref.empty_key_sentinel() : *found; | ||
| } | ||
| #endif |
Collaborator
There was a problem hiding this comment.
For context: This is the section I've been focussing on in my investigation.
Collaborator
|
I tested another H100 HBM3 node with CTK 12.3 and the performance regression is still present although much less pronounced (around 5-6% less throughput). The instruction diff between the baseline and this version is straightforward: if (tile.thread_rank() == 0) {...}compiles to which sets The Compared to the baseline, the new version additionally computes the mask of participating threads ( I have some profiles which I will share via Slack since I had to use an internal toolchain. |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
This PR updates new open addressing implementations to use
cg::invoke_onewhen possible.It doesn't change legacy implementations like multimap or dynamic map, etc.