Add fixed-size segment env overloads for DeviceSegmentedReduce by gonidelis · Pull Request #8097 · NVIDIA/cccl

gonidelis · 2026-03-18T23:17:40Z

Split 2/2 for env SegmentedReduce

Adds fixed-size overloads. Ideally would be merged after Add env SegmentedReduce (non fixed-size overloads) #7795
Merge first: Add public operator<<(ostream, cub::KeyValuePair) #8335

closes #7550

copy-pr-bot · 2026-03-18T23:17:43Z

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.

bernhardmgruber

First pass through the header

bernhardmgruber

I pushed a bit of refactoring so you can print and compare vectors of pairs.

bernhardmgruber · 2026-03-26T08:37:18Z

 {
 private:
+  template <typename InputIteratorT, typename OutputIteratorT>
+  CUB_RUNTIME_FUNCTION static cudaError_t fixed_size_arg_min_dispatch(


Great work! Now try to unify this function with fixed_size_arg_max_dispatch. There are two different tokens that matter ;)

bernhardmgruber · 2026-03-26T08:39:44Z

-  //! - Provides "run-to-run" determinism for pseudo-associative reduction
-  //!   (e.g., addition of floating point types) on the same GPU device.
-  //!   However, results for pseudo-associative reduction may be inconsistent
-  //!   from one device to a another device of a different compute-capability
-  //!   because CUB can employ different tile-sizing for different architectures.


Q: Why is this comment deleted? Doesn't the fixed-size segmented reduction provide run-to-run determinism? Was the comment wrong?

this is not an env overload

bernhardmgruber · 2026-03-26T09:22:56Z

+  cudaStream_t custom_stream;
+  REQUIRE(cudaSuccess == cudaStreamCreate(&custom_stream));


Important: Please use cuda::stream instead of manual resource management

- Retain offset_t/SegmentSizeT comments across all fixed-size env overloads - Retain PR NVIDIA#414 breaking-change comments in ArgMin/ArgMax env overloads - Refactor fixed-size ArgMin/ArgMax into shared private helpers (fixed_size_arg_min_dispatch, fixed_size_arg_max_dispatch) to reduce duplication between env and non-env overloads

…) but outside of the literalinclude guards

copy-pr-bot · 2026-04-13T13:44:23Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

bernhardmgruber

Two more things, otherwise LGTM

bernhardmgruber · 2026-04-13T21:11:27Z

+          // allow serializing only pairs with arithmetic members otherwise normal_iterator tests complain
+          enable_if_t<::cuda::std::is_arithmetic_v<T1>&& ::cuda::std::is_arithmetic_v<T2>, int> = 0>


I don't understand why this is necessary. If operator<< is called for a pair of iterators, and you can't serialize an iterator, then why does removing this overload allow the program to compile? what other operator<< is chosen instead?

Which test fails to compile if you remove the enable_if?

ci was failing in libcudacxx/test/libcudacxx/std/algorithms/alg.modifying/alg.partitions/pstl_partition_copy.cu in L#67 before i add this sfinae

Thx! I understand now what's happening. Catch2 was trying to find an operator<<(osteam, pair<some thrust iterators>). Before, it didn't fine one and just did not generate code for printing the value. With the above change (and no constraint) it found one, but then the implementation of operator<< did not compile.

The trick here is, as you noticed, to make the operator<< SFINAE friendly. That means, that the operator<< will SFINAE itself out if it knows that its implementation would not compile. This way, a library like Catch2 can detect whether it can use the operator<< (by testing whether a call to that operator<< would compile).

I refined the constraint to use:

template <typename T1, typename T2, ::cuda::std::void_t<decltype(::cuda::std::declval<::std::ostream>() << ::cuda::std::declval<T1>() << ::cuda::std::declval<T2>())>* = nullptr> ::std::ostream& operator<<(::std::ostream& os, const pair<T1, T2>& pair)

This tests whether T1 and T2 can we written to an ostream as part of the interface of operator<<. If they can't be written, the operator<< is SFINAE-ed out.

github-actions · 2026-04-14T09:58:31Z

🥳 CI Workflow Results

🟩 Finished in 1h 10m: Pass: 100%/269 | Total: 2d 00h | Max: 45m 54s | Hits: 98%/177147

See results here.

…A#8097) * Add fixed-size segment env overloads for DeviceSegmentedReduce * Fix critical issues plus<> specilization * feedback address - Retain offset_t/SegmentSizeT comments across all fixed-size env overloads - Retain PR NVIDIA#414 breaking-change comments in ArgMin/ArgMax env overloads - Refactor fixed-size ArgMin/ArgMax into shared private helpers (fixed_size_arg_min_dispatch, fixed_size_arg_max_dispatch) to reduce duplication between env and non-env overloads * Refactor * More refactoring * Refactor: unify fixed-size ArgMin/ArgMax dispatch into a single function * Make using offset_t = int explicit * Substitute cudaStream_t with cuda::stream and use always stream.sync() but outside of the literalinclude guards * Guard Sum with sfinae * Sfinae operator<<(os,...) for cuda::std::pairs * Do not specialize std::plus<> --------- Co-authored-by: Bernhard Manfred Gruber <bernhardmgruber@gmail.com>

github-project-automation Bot added this to CCCL Mar 18, 2026

github-project-automation Bot moved this to Todo in CCCL Mar 18, 2026

cccl-authenticator-app Bot moved this from Todo to In Progress in CCCL Mar 18, 2026

gonidelis mentioned this pull request Mar 18, 2026

Add env SegmentedReduce (non fixed-size overloads) #7795

Merged

gonidelis marked this pull request as ready for review March 18, 2026 23:27

gonidelis requested a review from a team as a code owner March 18, 2026 23:27

gonidelis requested a review from elstehle March 18, 2026 23:27

cccl-authenticator-app Bot moved this from In Progress to In Review in CCCL Mar 18, 2026

This comment has been minimized.

Sign in to view

gonidelis force-pushed the seg_red_fixedsize_env branch from b728f61 to 5f48a59 Compare March 19, 2026 02:35

This comment has been minimized.

Sign in to view

gonidelis requested a review from bernhardmgruber March 24, 2026 15:18

bernhardmgruber reviewed Mar 25, 2026

View reviewed changes

gonidelis force-pushed the seg_red_fixedsize_env branch from 8c9d411 to 047a10b Compare March 26, 2026 00:15

This comment has been minimized.

Sign in to view

bernhardmgruber requested a review from a team as a code owner March 26, 2026 09:21

bernhardmgruber reviewed Mar 26, 2026

View reviewed changes

This comment has been minimized.

Sign in to view

gonidelis requested review from a team as code owners April 8, 2026 19:38

gonidelis requested review from alliepiper, caugonnet and kkraus14 April 8, 2026 19:38

bernhardmgruber force-pushed the seg_red_fixedsize_env branch from 3d76549 to 0acae2d Compare April 9, 2026 08:02

This comment has been minimized.

Sign in to view

gonidelis and others added 9 commits April 13, 2026 02:57

Add fixed-size segment env overloads for DeviceSegmentedReduce

700eaef

Fix critical issues plus<> specilization

b0784a8

Refactor

0ade842

More refactoring

12874fa

Refactor: unify fixed-size ArgMin/ArgMax dispatch into a single function

091662c

Make using offset_t = int explicit

daa85bf

Substitute cudaStream_t with cuda::stream and use always stream.sync(…

e65cf05

…) but outside of the literalinclude guards

Guard Sum with sfinae

f82642c

gonidelis force-pushed the seg_red_fixedsize_env branch from 0acae2d to f82642c Compare April 13, 2026 10:00

gonidelis requested a review from bernhardmgruber April 13, 2026 10:01

This comment has been minimized.

Sign in to view

bernhardmgruber reviewed Apr 13, 2026

View reviewed changes

gonidelis force-pushed the seg_red_fixedsize_env branch from 44d2941 to cee2dba Compare April 13, 2026 23:55

gonidelis added 2 commits April 13, 2026 16:57

Sfinae operator<<(os,...) for cuda::std::pairs

ea6cac4

Do not specialize std::plus<>

3c44079

gonidelis force-pushed the seg_red_fixedsize_env branch from cee2dba to 3c44079 Compare April 13, 2026 23:57

gonidelis enabled auto-merge (squash) April 14, 2026 00:57

This comment has been minimized.

Sign in to view

bernhardmgruber approved these changes Apr 14, 2026

View reviewed changes

bernhardmgruber added 2 commits April 14, 2026 09:42

Refine operator<<(os, pair<T, U) constraint

548f40e

Fix clang-tidy

c7489c8

gonidelis merged commit fb67b5d into NVIDIA:main Apr 14, 2026
285 of 288 checks passed

github-project-automation Bot moved this from In Review to Done in CCCL Apr 14, 2026

bernhardmgruber mentioned this pull request Apr 14, 2026

simplify dispatch segmented reduce to use latest dispatch and new tunings API #8332

Merged

3 tasks done

		cudaStream_t custom_stream;
		REQUIRE(cudaSuccess == cudaStreamCreate(&custom_stream));

		// allow serializing only pairs with arithmetic members otherwise normal_iterator tests complain
		enable_if_t<::cuda::std::is_arithmetic_v<T1>&& ::cuda::std::is_arithmetic_v<T2>, int> = 0>

Conversation

gonidelis commented Mar 18, 2026 • edited by bernhardmgruber Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

copy-pr-bot Bot commented Mar 18, 2026

Uh oh!

This comment has been minimized.

This comment has been minimized.

bernhardmgruber left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

This comment has been minimized.

bernhardmgruber left a comment

Choose a reason for hiding this comment

Uh oh!

bernhardmgruber Mar 26, 2026

Choose a reason for hiding this comment

Uh oh!

bernhardmgruber Mar 26, 2026

Choose a reason for hiding this comment

Uh oh!

gonidelis Apr 8, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

bernhardmgruber Mar 26, 2026

Choose a reason for hiding this comment

Uh oh!

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

copy-pr-bot Bot commented Apr 13, 2026

Uh oh!

bernhardmgruber left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

bernhardmgruber Apr 13, 2026

Choose a reason for hiding this comment

Uh oh!

gonidelis Apr 13, 2026

Choose a reason for hiding this comment

Uh oh!

bernhardmgruber Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

This comment has been minimized.

github-actions Bot commented Apr 14, 2026

🥳 CI Workflow Results

🟩 Finished in 1h 10m: Pass: 100%/269 | Total: 2d 00h | Max: 45m 54s | Hits: 98%/177147

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

gonidelis commented Mar 18, 2026 •

edited by bernhardmgruber

Loading