[CUB] Use BlockLoadToShared in DeviceMerge
#6077
Merged
+267
−57
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
Contributor
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
3 tasks
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as outdated.
This comment was marked as outdated.
pauleonix
force-pushed
the
device_merge_tma
branch
from
e05ad2b to
582da7c
Compare
This comment was marked as resolved.
This comment was marked as resolved.
Contributor
Author
|
B200 (UBLKCPY)['/home/pgrossebley/SM_100_merge_keys_final_old.json', '/home/pgrossebley/SM_100_merge_keys_final_newest.json'] base[0] NVIDIA B200
Summary
H200 (UBLKCPY)['/home/pgrossebley/SM_90_merge_keys_final_old.json', '/home/pgrossebley/SM_90_merge_keys_final_newest.json'] base[0] NVIDIA H200
Summary
A100 (LDGSTS)['/home/pgrossebley/SM_80_merge_keys_final_old.json', '/home/pgrossebley/SM_80_merge_keys_final_newest.json'] base[0] NVIDIA A100 80GB PCIe
Summary
RTX 5090 (UBLKCPY)['/home/pgrossebley/SM_120_merge_keys_old_final_CTK13.json', '/home/pgrossebley/SM_120_merge_keys_newest_final_CTK13.json'] base[0] NVIDIA GeForce RTX 5090
Summary
|
Contributor
Author
|
B200 (UBLKCPY)['/home/pgrossebley/SM_100_merge_pairs_final_old.json', '/home/pgrossebley/SM_100_merge_pairs_final_newest.json'] base[0] NVIDIA B200
Summary
H200 (UBLKCPY)['/home/pgrossebley/SM_90_merge_pairs_final_old.json', '/home/pgrossebley/SM_90_merge_pairs_final_newest.json'] base[0] NVIDIA H200
Summary
A100 (LDGSTS)Did not fit due to GH comment character limit, see newer comment below. RTX 5090 (UBLKCPY)['/home/pgrossebley/SM_120_merge_pairs_old_final_CTK13.json', '/home/pgrossebley/SM_120_merge_pairs_newest_final_CTK13.json'] base[0] NVIDIA GeForce RTX 5090
Summary
|
This comment was marked as outdated.
This comment was marked as outdated.
Contributor
|
The benchmark looks very promising! There are few runs though that regressed a lot, like some 2^16 workloads with more than 20% slowdown. I think it would be worthwhile to understand what's going on there. |
pauleonix
force-pushed
the
device_merge_tma
branch
from
582da7c to
14a356a
Compare
This comment was marked as resolved.
This comment was marked as resolved.
This comment has been minimized.
This comment has been minimized.
Contributor
Author
|
A100 (LDGSTS)['/home/pgrossebley/SM_80_merge_pairs_final_old.json', '/home/pgrossebley/SM_80_merge_pairs_final_newest.json'] base[0] NVIDIA A100 80GB PCIe
Summary
|
Contributor
Author
|
@bernhardmgruber It seems that benchmarking small problem sizes (2^16) is not super reproducible on the RTX 5090. I replaced the old results with a newer run which has far less and smaller slowdowns. The few bigger slowdowns at 2^16 elements appear for different benchmarks than the slowdowns in the previous results. I can also see differences in the 2^16 results between running just those benchmarks or the full default set including bigger sizes. The new run also has a slightly different environment hopefully nearer to the cluster (CTK 12.9->13.0, graphical->multiuser and some other probably less important things like kernel updates, turned on persistence mode) which might have somewhat increased reproducibility (runtimes for the old implementation have profited more than runtimes for the new implementation). |
bernhardmgruber
added a commit
to bernhardmgruber/cccl
that referenced
this pull request
Oct 7, 2025
This avoids needing to specify the cache modifier in two places and eases integration of NVIDIA#6077
bernhardmgruber
added a commit
that referenced
this pull request
Oct 7, 2025
This avoids needing to specify the cache modifier in two places and eases integration of #6077
bernhardmgruber
force-pushed
the
device_merge_tma
branch
from
14a356a to
cf6dfa4
Compare
Contributor
Contributor
|
/ok to test cf6dfa4 |
Contributor
|
I rebased the branch on main. |
cub/cub/agent/agent_merge.cuh
Outdated
| (void) translate_indices; | ||
|
|
||
| item_type* items1_shared; | ||
| if constexpr (keys_use_block_load_to_shared) |
Contributor
There was a problem hiding this comment.
Important: I think this should be items_use_block_load_to_shared
Contributor
There was a problem hiding this comment.
Yes, this is exposed by new tests: #6455. But only after some refactoring that I have not yet pushed to this branch.
bernhardmgruber
force-pushed
the
device_merge_tma
branch
from
cf6dfa4 to
0d89c9f
Compare
bernhardmgruber
force-pushed
the
device_merge_tma
branch
from
fe08eaf to
c2c52f0
Compare
Contributor
|
/ok to test c2c52f0 |
Contributor
|
@pauleonix and @elstehle #6460 offers a refactoring and additional tuning policies on top of this PR. We should merge it into this PR and then merge this PR. |
This reverts commit a9fd762.
Contributor
🥳 CI Workflow Results🟩 Finished in 5h 21m: Pass: 100%/81 | Total: 4d 21h | Max: 5h 20m | Hits: 50%/72769See results here. |
bernhardmgruber
approved these changes
Nov 6, 2025
Contributor
|
Great work @pauleonix! The first CUB algo is now using |
Contributor
|
That is awesome @pauleonix , great job 🎉 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
Description
closes #6005 (previous PR #5926 was wrongly closed, go there for slightly outdated performance results until I generate new ones)
Since
DeviceMergeneeds its data in shared memory either way, it is a good candidate to demonstrate the newBlockLoadToSharedinterface in the real world. It also comes with some complications since we know the size of the output tile at compile-time, but how many elements to load from each of the ranges to be merged (calculated at runtime using merge-path). This can be solved by introducing additional padding (size known at compile-time) for in between the two dynamically sized input buffers.Checklist