[cub] Replace assert with CHECK or REQUIRE in tests

[davebayer]

Using assert outside of libcu++ for testing is not safe, because we compile with optimizations enabled and that meanst NDEBUG is defined, thus assert expands to noop. This PR replaces them with catch2 test macros.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 070b0c74-4bea-4103-8422-4a110331012d

📥 Commits

Reviewing files that changed from the base of the PR and between 329af26 and c9b107a.

📒 Files selected for processing (9)

• cub/cub/detail/integer_utils.cuh
• cub/test/catch2_test_block_load_to_shared.cu
• cub/test/catch2_test_block_radix_sort_custom.cu
• cub/test/catch2_test_device_decoupled_look_back.cu
• cub/test/catch2_test_device_segmented_scan.cu
• cub/test/catch2_test_warp_scan.cu
• cub/test/catch2_test_warp_scan_api.cu
• cub/test/catch2_test_warp_scan_partial_helper.cuh
• cub/test/internal/catch2_test_integer_utils.cu

💤 Files with no reviewable changes (2)

• cub/test/internal/catch2_test_integer_utils.cu
• cub/cub/detail/integer_utils.cuh

✅ Files skipped from review due to trivial changes (1)

• cub/test/catch2_test_warp_scan.cu

---

📝 Walkthrough

Summary by CodeRabbit

• Tests
  • Replaced C runtime asserts with Catch2 CHECK/REQUIRE across multiple device and host tests for clearer test failures and diagnostics.
• Chores
  • Removed internal integer/floating-point utility helpers from the codebase.

important:

Walkthrough

Test-only changes: device- and host-side C assert() usages are replaced with Catch2 test macros (REQUIRE, REQUIRE_DEVICE, CHECK) across multiple CUDA test files; a private integer utility header cub/cub/detail/integer_utils.cuh was removed.

important:

Changes

Catch2 assertion framework migration

Layer / File(s)	Summary
Include setup and precondition assertions cub/test/catch2_test_block_radix_sort_custom.cu, cub/test/catch2_test_device_decoupled_look_back.cu, cub/test/catch2_test_device_segmented_scan.cu, cub/test/catch2_test_warp_scan_partial_helper.cuh, cub/test/catch2_test_warp_scan.cu	Top-of-file includes remove NDEBUG/<cassert> patterns where present and precondition validations (e.g., small_size > 0, host-result size divisibility) use Catch2 REQUIRE.
Block-level kernels: alignment and radix-sort checks cub/test/catch2_test_block_load_to_shared.cu, cub/test/catch2_test_block_radix_sort_custom.cu	Device-side shared-memory alignment check and block radix-sort kernel key/value verifications (ascending/descending, bit-subrange, blocked-to-striped, key/pair variants) use Catch2 REQUIRE_DEVICE/REQUIRE instead of assert.
Scan-related kernel checks cub/test/catch2_test_device_decoupled_look_back.cu, cub/test/catch2_test_warp_scan_api.cu	Decoupled look-back tile data/aggregate validations and warp-scan API custom functor operand-range checks use Catch2 REQUIRE_DEVICE/REQUIRE.
Integer/floating-point utility tests and header removal cub/test/internal/catch2_test_integer_utils.cu, cub/cub/detail/integer_utils.cuh	Integer split/merge and floating-point comparable-integer tests migrate assert -> Catch2 CHECK (fix assignment-in-assert to equality); the internal header cub/cub/detail/integer_utils.cuh was removed.

important: Possibly related PRs

• NVIDIA/cccl#8928: Introduces device-capable Catch2 macros (including REQUIRE_DEVICE) used by these test migrations.

important: Suggested reviewers

• pauleonix
• bernhardmgruber
• srinivasyadav18
• gevtushenko

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 2

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 0274cd35-7d0c-401a-865f-079699d7dda5

📥 Commits

Reviewing files that changed from the base of the PR and between db93bd1 and f6b8406.

📒 Files selected for processing (8)

• cub/test/catch2_test_block_load_to_shared.cu
• cub/test/catch2_test_block_radix_sort_custom.cu
• cub/test/catch2_test_device_decoupled_look_back.cu
• cub/test/catch2_test_device_segmented_scan.cu
• cub/test/catch2_test_warp_scan.cu
• cub/test/catch2_test_warp_scan_api.cu
• cub/test/catch2_test_warp_scan_partial_helper.cuh
• cub/test/internal/catch2_test_integer_utils.cu

[davebayer]

I've removed the source file & test. The functionality is not used anywhere in cub, plus it was never actually tested, because the assert expanded to noop instead of the test due to missing #undef NDEBUG

[davebayer]

@gevtushenko is strongly for using REQUIRE here and in the rest of the PR

[bernhardmgruber]

I remain against. As a person that has to figure out what's going on in case of a test failure, I need the largest context possible of meaningful information.

That means, I don't want to see dependent failures, like the failed comparison with an expected result after a failed call to radix sort. That call to radix sort needs a REQUIRE of course. But a comparison of the results keys and values needs a CHECK since in case of a mismatch in the result, I want to see the mismatch of the keys and values. This may allow me to deduce better, what kind of failure happened, without first modifying the test code and recompiling.

@gevtushenko what would be the advantage of REQUIRE here? I don't see it.

[pauleonix]

To add some context from the code review hour: According to @davebayer CHECK in device code is currently implemented to do the same as REQUIRE because anything else would be hard to implement. So with the current implementation you get REQUIRE semantics anyway and @gevtushenko just argued that this state is confusing and that CHECK should not compile in device code if it can't provide the correct semantics.

[bernhardmgruber]

I see. I didn't see this was device code. Fine then!

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 1h 11m: Pass: 100%/285 | Total: 2d 08h | Max: 42m 54s | Hits: 89%/217105

See results here.