[davebayer]

No description provided.

[coderabbitai]

Summary

Summary by CodeRabbit

• Bug Fixes
  • Improved group mapping compatibility with forward-only ranges.
• Tests
  • Expanded coverage for exhaustive and partial group mappings across runtime and compile-time ranges.

__group_as_map now retrieves unit counts by advancing an iterator. The tests add exhaustive and non-exhaustive mapping checks for forward-only ranges.

Changes: Forward-only group mapping

Layer / File(s)	Summary
Iterator-based count mapping cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh, cudax/test/coop/group/mapping/group_as.cu	__group_as_map reads unit counts sequentially through an iterator. Tests cover exhaustive and non-exhaustive mappings for {4, 12, 16}, and a non-exhaustive mapping for {4, 6, 8}.

Possibly related issues

• [FEA]: Add static unit count dispatch support for groups #12551: The iterator-based access is relevant to the issue’s compile-time-wrapped unit-count array, but the requested static-count dispatch is outside this change.
• [FEA]: Extend group_as to support compile-time known ranges #12554: The change adds compile-time-range tests and iterator-based access, but the evidence does not show support for every requested range form.

Possibly related PRs

• NVIDIA/cccl#12223: This PR changes group_as to accept viewable ranges; the current change replaces indexed count access with iterator-based access.

Change: Bug fix

Merge Risk: Low · up to 279c8

Certain supported custom count ranges may fail to compile or be read after exhaustion. These compatibility gaps need correction or owner awareness.

Security architecture risk: Low · up to 279c8

The intended forward-only inputs preserve the existing mapping behavior. A narrow contract concern remains: mapping traverses the counts repeatedly, but its input requirements do not enforce reusable iterators. No privilege expansion or isolation bypass is demonstrated.

Retained concerns

• Low · architecture · inferred: The widened runtime range acceptance does not express the multipass invariant required by mapping. Accumulation can consume a sized single-pass range before the new iterator traversal reads it, potentially producing invalid iterator access or incorrect mapping within the calling device kernel. This is an inferred failure-containment concern under the declared constraints; documented indexing requirements and the absence of a concrete single-pass consumer limit the exposure established here.

Security review details

Security Blast Radius

• inferred — The supported effect is on subgroup membership and lane-mask metadata in calling device kernels. The available evidence does not establish a downstream authorization, tenant-isolation, credential, or service boundary that consumes these results.

Trust Boundaries and Controls

• observed — The named TestKernel and operator() routing subjects are inside an anonymous namespace and are launched by a test case with fixed count fixtures. Their changes do not constitute a new externally exposed production entrypoint.

Hardening Proposals

• proposed — Make the reusable-iteration requirement explicit in the dynamic range contract, for example through an appropriate forward-range constraint, and align the documentation with the intended forward-only support.

- [ ] Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

[coderabbitai]

Actionable comments posted: 2

Review info

Run configuration

• Source mode: Auto
• Repository: NVIDIA/cccl
• Branch: main
• File: .coderabbit.yaml
• Commit: dd063b7
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 877c280f-a1df-4386-9cfd-116def47bd47

Commits

Reviewing files that changed from the base of the PR and between dd063b7 and 279c80b.

Files selected for processing: 2

• cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh
• cudax/test/coop/group/mapping/group_as.cu

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

[coderabbitai]

Minor · Stability & Availability · Quick win

Supported by static analysis

Script executed:

sed -n '1,240p' cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh
rg -n 'group_as|__unit_counts|_CCCL_ASSERT' cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh

Repository: NVIDIA/cccl

Length of output: 13483

---

Script executed:

git diff dd063b73e63e18f911182f79879fbf7d558ebe79 279c80bebbf42e46b24c012727c7a4a4e13b5f48 -- cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh
rg -n -F -e '#define _CCCL_ASSERT' -e '#  define _CCCL_ASSERT' -e '#    define _CCCL_ASSERT' libcudacxx/include
rg -n -F -e 'sized_range' -e 'forward_range' libcudacxx/include/cuda/std/__ranges/concepts.h
rg -n -F -e 'group_as' cudax/test/coop cudax/include/cuda/experimental/coop/__group/mapping 2>/dev/null

Repository: NVIDIA/cccl

Length of output: 15944

---

Script executed:

cat -n libcudacxx/include/cuda/std/__cccl/assert.h | sed -n '34,170p'
cat -n libcudacxx/include/cuda/std/__numeric/accumulate.h | sed -n '20,110p'
cat -n libcudacxx/include/cuda/std/__ranges/concepts.h | sed -n '55,120p'
cat -n libcudacxx/include/cuda/std/__ranges/all.h | sed -n '35,155p'
cat -n cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh | sed -n '53,153p'
cat -n cudax/test/coop/group/mapping/group_as.cu | sed -n '28,105p;325,350p'
rg -n -F -e 'group_as' docs/cudax/coop/group
rg -n -F -e 'weakly_incrementable' -e 'input_iterator' libcudacxx/include/cuda/std/__iterator/concepts.h

Repository: NVIDIA/cccl

Length of output: 27079

---

Script executed:

cat -n docs/cudax/coop/group/mapping/group_as.rst | sed -n '1,105p'
cat -n libcudacxx/include/cuda/std/__iterator/concepts.h | sed -n '92,158p'
cat -n libcudacxx/include/cuda/std/__ranges/ref_view.h | sed -n '35,150p'
cat -n libcudacxx/include/cuda/std/__ranges/access.h | sed -n '230,320p'

Repository: NVIDIA/cccl

Length of output: 13390

---

important: Process single-pass counts in one traversal.

With device assertions disabled, non-exhaustive validation can consume a single-pass range before the new traversal dereferences it. With assertions enabled, the earlier assertion can also consume the range and cause a valid mapping to return invalid. Combine validation and mapping in one traversal while preserving the existing range contract.

[coderabbitai]

Minor · Functional Correctness · Quick win

Supported by static analysis

Script executed:

sed -n '1,240p' cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh
rg -n 'group_as|sized_range|forward_range|input_range' cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh cudax/test/coop/group/mapping/group_as.cu

Repository: NVIDIA/cccl

Length of output: 20410

---

Script executed:

cat -n cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh | sed -n '48,155p;245,275p'
git diff dd063b73e63e18f911182f79879fbf7d558ebe79 279c80bebbf42e46b24c012727c7a4a4e13b5f48 -- cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh
rg -n -F -e 'weakly_incrementable' -e 'input_or_output_iterator' -e 'input_iterator' libcudacxx/include/cuda/std/__iterator/concepts.h
rg -n -F -e 'sized_range' -e 'viewable_range' -e 'input_range' libcudacxx/include/cuda/std/__ranges/concepts.h
rg -n -F -e 'accumulate(' libcudacxx/include/cuda/std/__numeric/accumulate.h
cat -n libcudacxx/include/cuda/std/__numeric/accumulate.h | sed -n '20,85p'
cat -n libcudacxx/include/cuda/std/__iterator/concepts.h | sed -n '70,200p'
cat -n libcudacxx/include/cuda/std/__ranges/concepts.h | sed -n '45,170p'
cat -n cudax/test/coop/group/mapping/group_as.cu | sed -n '125,207p'

Repository: NVIDIA/cccl

Length of output: 26280

---

important: Avoid dereferencing the postfix result.

A sized input range can have fresh iterators whose postfix increment returns void. For such a range, group_as::map can fail to compile. Read the value before using prefix increment.

Fix

--- "a/cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh"
+++ "b/cudax/include/cuda/experimental/coop/__group/mapping/group_as.cuh"
@@ -100,7 +100,8 @@
   _CCCL_PRAGMA_UNROLL_FULL()
   for (::cuda::std::uint32_t __i = 0; __i < __curr_ngroups; ++__i)
   {
-    const auto __i_count = static_cast<::cuda::std::uint32_t>(*__it++);
+    const auto __i_count = static_cast<::cuda::std::uint32_t>(*__it);
+    ++__it;
     if (__prev_unit_rank < __sum + __i_count)
     {
       const auto __group_rank = __prev_mapping_result.group_rank() * __curr_ngroups + __i;

Committable suggestion

IMPORTANT
Review the replacement and indentation, then test before committing.

Suggested change

	const auto __i_count = static_cast<::cuda::std::uint32_t>(*__it++);
	const auto __i_count = static_cast<::cuda::std::uint32_t>(*__it);
	++__it;