[c2h] Make catch macros work on device by davebayer · Pull Request #8928 · NVIDIA/cccl

davebayer · 2026-05-12T16:23:45Z

Catch 2 test macros are unsupported in device code, so we have other alternatives like CUDAX_REQUIRE or CCCLRT_REQUIRE that should work in both host and device code.

I don't like this, so I found a way to allow us using the original macros directly. If CATCH_CONFIG_PREFIX_ALL macro is defined, catch 2 prepends these macros with CATCH_ and the non-prefixed versions are not defined. We can then define those manually and define our own behaviour in device code.

The output for failed tests looks something like:

/home/dabayer/cccl/cudax/test/group/synchronizer/lane_synchronizer.cu:48:
   CHECK(synchronizer_instance.__lane_mask_ != (this_lane_mask | another_lane_mask)) failed
   block [0, 0, 0], thread [4, 3, 0]

I've also implemented some extensions, such as:

REQUIRE_CUDA(cuSomeDriverCall(...));
CHECK_CUDA(cuSomeDriverCall(...));

REQUIRE_CUDART(cudaSomeCudaRuntimeCall(...));
CHECK_CUDART(cudaSomeCudaRuntimeCall(...));

davebayer · 2026-05-12T20:37:52Z

+#define CUDAX_REQUIRE(condition) REQUIRE(condition)

-#define CUDAX_CHECK(condition)                                                                             \
-  NV_IF_ELSE_TARGET(NV_IS_DEVICE,                                                                          \
-                    (cudax_require_impl(condition, #condition, __FILE__, __LINE__, __PRETTY_FUNCTION__);), \
-                    (CHECK(condition);))
+#define CUDAX_CHECK(condition) CHECK(condition)

-#define CUDAX_FAIL(message) /*                                                                   */ \
-  NV_IF_ELSE_TARGET(NV_IS_DEVICE, /*                                                             */ \
-                    (cudax_require_impl(false, message, __FILE__, __LINE__, __PRETTY_FUNCTION__);), \
-                    (FAIL(message);))
+#define CUDAX_FAIL(message) FAIL(message)

-#define CUDAX_CHECK_FALSE(condition) CUDAX_CHECK(!(condition))
+#define CUDAX_CHECK_FALSE(condition) CHECK_FALSE(condition)


I will remove these in a separate PR

davebayer · 2026-05-12T20:38:02Z

 #define CUDART(call) REQUIRE((call) == cudaSuccess)

-// There is a problem with clang-cuda and nv/target, but we don't need the device side macros yet,
-// disable them for now
-#if _CCCL_CUDA_COMPILER(CLANG)
-#  define CCCLRT_REQUIRE(condition)     REQUIRE(condition)
-#  define CCCLRT_CHECK(condition)       CHECK(condition)
-#  define CCCLRT_FAIL(message)          FAIL(message)
-#  define CCCLRT_CHECK_FALSE(condition) CCCLRT_CHECK(!(condition))
-
-#else // _CCCL_CUDA_COMPILER(CLANG)
-#  define CCCLRT_REQUIRE(condition)                                                                           \
-    NV_IF_ELSE_TARGET(NV_IS_DEVICE,                                                                           \
-                      (ccclrt_require_impl(condition, #condition, __FILE__, __LINE__, __PRETTY_FUNCTION__);), \
-                      (REQUIRE(condition);))
-
-#  define CCCLRT_CHECK(condition)                                                                             \
-    NV_IF_ELSE_TARGET(NV_IS_DEVICE,                                                                           \
-                      (ccclrt_require_impl(condition, #condition, __FILE__, __LINE__, __PRETTY_FUNCTION__);), \
-                      (CHECK(condition);))
-
-#  define CCCLRT_FAIL(message) /*                                                                   */ \
-    NV_IF_ELSE_TARGET(NV_IS_DEVICE, /*                                                             */  \
-                      (ccclrt_require_impl(false, message, __FILE__, __LINE__, __PRETTY_FUNCTION__);), \
-                      (FAIL(message);))
-
-#  define CCCLRT_CHECK_FALSE(condition) CCCLRT_CHECK(!(condition))
-#endif // _CCCL_CUDA_COMPILER(CLANG)
+#define CCCLRT_REQUIRE(condition) REQUIRE(condition)
+
+#define CCCLRT_CHECK(condition) CHECK(condition)
+
+#define CCCLRT_FAIL(message) FAIL(message)
+
+#define CCCLRT_CHECK_FALSE(condition) CHECK_FALSE(condition)

 // Explicit device side require macros for clang-cuda
-#define CCCLRT_REQUIRE_DEVICE(condition) \
-  ccclrt_require_impl(condition, #condition, __FILE__, __LINE__, __PRETTY_FUNCTION__);
-#define CCCLRT_CHECK_DEVICE(condition) \
-  ccclrt_require_impl(condition, #condition, __FILE__, __LINE__, __PRETTY_FUNCTION__);
-#define CCCLRT_FAIL_DEVICE(message)          ccclrt_require_impl(false, message, __FILE__, __LINE__, __PRETTY_FUNCTION__);
-#define CCCLRT_CHECK_FALSE_DEVICE(condition) CCCLRT_CHECK_DEVICE(!(condition))
+#define CCCLRT_REQUIRE_DEVICE(condition)     REQUIRE_DEVICE(condition)
+#define CCCLRT_CHECK_DEVICE(condition)       CHECK(condition)
+#define CCCLRT_FAIL_DEVICE(message)          FAIL(message)
+#define CCCLRT_CHECK_FALSE_DEVICE(condition) CHECK_FALSE(condition)


davebayer · 2026-05-12T20:38:31Z

 // example-end custom-type

-static __host__ std::ostream& operator<<(std::ostream& os, const custom_t& self)
+__host__ std::ostream& operator<<(std::ostream& os, const custom_t& self)


I got warnings about these functions being unused 🤷

IMO then we should remove them if they really are unused.

They are used, just not during device pass and that makes cicc complain

coderabbitai · 2026-05-14T06:42:46Z

📝 Walkthrough

Summary by CodeRabbit

New Features
- Added a unified Catch2-style testing layer with device-side assertion support and expanded test helper macros.
Refactor
- Simplified and consolidated test assertion macros to consistently map host/device behavior.
- Test integration updated to use the new helper layer and a compile-time definition that affects test builds.
Bug Fixes
- Fixed test linkage and assertion usage to resolve compilation/linkage issues and adjusted several test assertions for correctness.

important:

Walkthrough

Adds a host/device-compatible Catch2 macro header that implements device-side assertion behavior and aliases, exposes REQUIRE/CHECK/FAIL/TEMPLATE/BDD/matcher macros for host and device, and wires the build and helper macros to use prefixed Catch2 forms.

important:

Changes

Catch2 host+device testing infrastructure

Layer / File(s)	Summary
Core C2H Catch2 macro foundation `c2h/include/c2h/catch2_test_macros.h`	New header providing host/device-aware assertion and test macros (REQUIRE/CHECK/FAIL, TEST_CASE/SECTION, BDD, TEMPLATE/TEMPLATE_LIST, matcher/throws forms), device failure printer that logs file/line and CUDA indices and invokes `::__trap()`, plus `REQUIRE_DEVICE` and CUDA helper macros.
C2H build integration and helper wiring `c2h/CMakeLists.txt`, `c2h/include/c2h/catch2_test_helper.h`	CMake target `cccl.c2h` adds public compile definition `CATCH_CONFIG_PREFIX_ALL`. Helper header now includes `c2h/catch2_test_macros.h` and updates `C2H_TEST_`/`C2H_TEST_LIST_` macros to use `CATCH_TEMPLATE_LIST_TEST_CASE_METHOD(...)`, wiring `::detail::nvtx_fixture` for non-fixture variants and passing `FIXTURE` for fixture variants.
Assertion macro simplification in test frameworks `cudax/test/common/testing.cuh`, `libcudacxx/test/libcudacxx/cuda/ccclrt/common/testing.cuh`	Replaced prior NV_IF_ELSE_TARGET/ccclrt/cudax dispatch with direct Catch2 macros: `CUDAX_` and `CCCLRT_` now map to `REQUIRE/CHECK/FAIL/CHECK_FALSE` (and `CCCLRT_REQUIRE_DEVICE` → `REQUIRE_DEVICE`); removed earlier conditional/device-implementation indirection.
Test header migrations and small test edits `cub/test/insert_nested_NVTX_range_guard.h`, `cudax/test/utility/optionally_static.cu`, `cudax/test/cufile/driver_register.cu`, `libcudacxx/test/libcudacxx/cuda/containers/buffer/access.cu`, `libcudacxx/test/libcudacxx/cuda/memory_resource/resources/shared_memory_pools.cu`	Tests switched includes to the new C2H helper/macros where applicable. Minor edits: void-cast of expressions inside `CHECK_THROWS_AS` and added `// NOLINT` on an intentional copy construction.
Operator linkage adjustments in tests `cub/test/catch2_test_device_radix_sort_custom.cu`, `cub/test/catch2_test_device_topk_api.cu`	Removed `static` from non-member `custom_t` operator overloads (`operator<<`, `operator==`, `operator<`, `operator>`), changing linkage to external within the translation units; implementations unchanged.

Suggested labels

cudax

Suggested reviewers

srinivasyadav18

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

coderabbitai

Actionable comments posted: 3

🧹 Nitpick comments (4)

c2h/include/c2h/catch2_test_macros.h (3)

27-37: 💤 Low value

suggestion: The internal macro C2H_INTERNAL_DEVICE_TEST_PRINT should use a reserved identifier pattern (e.g., _C2H_INTERNAL_DEVICE_TEST_PRINT) per libcudacxx coding style, which requires non-public symbols to use reserved identifiers (_ prefix for macros).

64-69: ⚡ Quick win

suggestion: Line 67 concatenates #EXPR #TYPE without spacing, producing output like "buf.at(4)std::out_of_range". Insert a space: `#EXPR " " `#TYPE or #EXPR ", " #TYPE``.

97-101: 💤 Low value

suggestion: Similar to CHECK_THROWS, line 100 concatenates #EXPR #TYPE without spacing. Use `#EXPR " " `#TYPE or #EXPR ", " #TYPE`` for readability.

libcudacxx/test/libcudacxx/cuda/ccclrt/common/testing.cuh (1)

37-40: ⚡ Quick win

suggestion: CCCLRT_CHECK_DEVICE and CCCLRT_FAIL_DEVICE currently forward to CHECK/FAIL instead of explicit device variants. This makes *_DEVICE behavior depend on non-device alias semantics under different toolchain/config combinations. Consider forwarding to CHECK_DEVICE and FAIL_DEVICE for consistency with CCCLRT_REQUIRE_DEVICE.

As per coding guidelines, this path should prioritize host/device availability and portability across supported CUDA toolchains.

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 6cfc2c30-30b1-4c59-8b4f-b0653a9160a2

📥 Commits

Reviewing files that changed from the base of the PR and between 5d79bc2 and bd9a3ea.

📒 Files selected for processing (11)

c2h/CMakeLists.txt
c2h/include/c2h/catch2_test_helper.h
c2h/include/c2h/catch2_test_macros.h
cub/test/catch2_test_device_radix_sort_custom.cu
cub/test/catch2_test_device_topk_api.cu
cub/test/insert_nested_NVTX_range_guard.h
cudax/test/common/testing.cuh
cudax/test/cufile/driver_register.cu
cudax/test/utility/optionally_static.cu
libcudacxx/test/libcudacxx/cuda/ccclrt/common/testing.cuh
libcudacxx/test/libcudacxx/cuda/containers/buffer/access.cu

Jacobfaib · 2026-05-14T14:30:03Z

+// <catch2/matchers/catch_matchers.hpp>
+
+#define REQUIRE_THROWS_WITH(...)    CATCH_REQUIRE_THROWS_WITH(__VA_ARGS__)
+#define REQUIRE_THROWS_MATCHES(...) CATCH_REQUIRE_THROWS_MATCHES(__VA_ARGS__)


In my humble (and correct!) opinion, REQUIRE_THROWS_MATCHES() is the only proper way to check exceptions thrown as it checks both the type and expected message so there is no ambiguity on which exception is being captured.

Perhaps we can take this time to fix up some defaults and just delete the other macros.

We definitely can! Let's move this discussion to slack/new issue. This PR really only aims to provide host + device working alternatives

I would strongly advise to stay with the exact semantics of Catch2. If REQUIRE_THROWS_MATCHES is indeed better, we should just use an agent to rewrite the tests not using it.

If we change the meaning of macros between C2H and Catch2, it will create a lot of confusion for maintainers.

Jacobfaib · 2026-05-14T14:32:08Z

+                    }))
+#define REQUIRE_NOTHROW(...) NV_IF_ELSE_TARGET(NV_IS_HOST, (CATCH_REQUIRE_NOTHROW(__VA_ARGS__);), (__VA_ARGS__;))
+
+#define CHECK(...)                                                             \


Similarly, we should consider removing CHECK() and replacing all instances with REQUIRE(). I am struggling to think of a good scenario where you would want test execution to continue after initial failure.

I think there could be value in having more information in the log from some external testing like CI or QA, so you might not have to reproduce the issue yourself, while REQUIRE would not give you enough information to figure out the failure. I wouldn't say its common and I can see using CHECK too often lead to a more messy log with multiple lines of repeated same failure for example. So its very case by case, but I think there is space for both of these macros

Agreed! But maybe someone else would have a valid use case for that

o you might not have to reproduce the issue yourself, while REQUIRE would not give you enough information to figure out the failure.

Ehhh, overall I think this ends up being rare in practice.

I see it similarly to C++ compiler error messages. Yes, once in a blue moon seeing the 30 different overloads of operator<< that GCC failed to match (and Very Helpful(TM) description for each on why they didn't match) might help you debug the failure, but usually all you want to see is that you accidentally typed bool is_less = foo << bar when you meant bool is_less = foo < bar, and that's the first error message in that whole stack.

More often than not, CHECK() leads to red herring errors, e.g. when you have

CHECK(container.size() >= n); // a lot of other unrelated code... CHECK(container[n] == foo); // oops

And now you are left trying parse the crash log instead.

I use CHECK() a lot where I want to validate a bunch of values after I ran a test. For example, I do a SortPairs and afterwards validate if the keys and values correspond to two reference arrays. I want to see both failures if they mismatch. Please keep the CHECK() macro.

Jacobfaib · 2026-05-14T14:33:33Z

 // example-end custom-type

-static __host__ std::ostream& operator<<(std::ostream& os, const custom_t& self)
+__host__ std::ostream& operator<<(std::ostream& os, const custom_t& self)


IMO then we should remove them if they really are unused.

pciolkosz · 2026-05-14T18:17:34Z

+// <catch2/catch2_test_macros.hpp>
+
+#define REQUIRE(...)                                                             \
+  NV_IF_ELSE_TARGET(NV_IS_HOST, (CATCH_REQUIRE(__VA_ARGS__);), ({                \


Don't you need to handle clang-coda here?

It seems to be working correctly in most cases, for other problematic cases I added REQUIRE_DEVICE

alliepiper

CMake signoff

bernhardmgruber

I am fine with this, but I want to point out that this may create additional maintenance work when upgrading Catch2.

bernhardmgruber · 2026-05-14T20:25:16Z

+                    }))
+#define REQUIRE_NOTHROW(...) NV_IF_ELSE_TARGET(NV_IS_HOST, (CATCH_REQUIRE_NOTHROW(__VA_ARGS__);), (__VA_ARGS__;))
+
+#define CHECK(...)                                                             \


I use CHECK() a lot where I want to validate a bunch of values after I ran a test. For example, I do a SortPairs and afterwards validate if the keys and values correspond to two reference arrays. I want to see both failures if they mismatch. Please keep the CHECK() macro.

bernhardmgruber · 2026-05-14T20:27:04Z

+// <catch2/matchers/catch_matchers.hpp>
+
+#define REQUIRE_THROWS_WITH(...)    CATCH_REQUIRE_THROWS_WITH(__VA_ARGS__)
+#define REQUIRE_THROWS_MATCHES(...) CATCH_REQUIRE_THROWS_MATCHES(__VA_ARGS__)


I would strongly advise to stay with the exact semantics of Catch2. If REQUIRE_THROWS_MATCHES is indeed better, we should just use an agent to rewrite the tests not using it.

If we change the meaning of macros between C2H and Catch2, it will create a lot of confusion for maintainers.

coderabbitai

Actionable comments posted: 1

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 6c5fe465-04e8-41f4-9cc6-a7dccf465957

📥 Commits

Reviewing files that changed from the base of the PR and between 694ebaf and 490c531.

📒 Files selected for processing (1)

c2h/include/c2h/catch2_test_macros.h

github-actions · 2026-05-14T23:16:39Z

🥳 CI Workflow Results

🟩 Finished in 2h 32m: Pass: 100%/358 | Total: 8d 09h | Max: 2h 31m | Hits: 68%/684690

See results here.

github-project-automation Bot added this to CCCL May 12, 2026

github-project-automation Bot moved this to Todo in CCCL May 12, 2026

cccl-authenticator-app Bot moved this from Todo to In Progress in CCCL May 12, 2026

davebayer force-pushed the c2h_test_macros branch 3 times, most recently from c5fc788 to 7da5c27 Compare May 12, 2026 17:23

davebayer commented May 12, 2026

View reviewed changes

davebayer marked this pull request as ready for review May 12, 2026 20:38

davebayer requested review from a team as code owners May 12, 2026 20:38

davebayer requested review from caugonnet, srinivasyadav18 and wmaxey May 12, 2026 20:38

cccl-authenticator-app Bot moved this from In Progress to In Review in CCCL May 12, 2026

This comment has been minimized.

Sign in to view

NVIDIA deleted a comment from copy-pr-bot Bot May 13, 2026

coderabbitai Bot reviewed May 14, 2026

View reviewed changes

Comment thread c2h/include/c2h/catch2_test_macros.h Outdated

Comment thread c2h/include/c2h/catch2_test_macros.h Outdated

Comment thread c2h/include/c2h/catch2_test_macros.h Outdated

This comment has been minimized.

Sign in to view

Jacobfaib reviewed May 14, 2026

View reviewed changes

pciolkosz reviewed May 14, 2026

View reviewed changes

pciolkosz approved these changes May 14, 2026

View reviewed changes

davebayer added 5 commits May 14, 2026 20:34

[c2h] Make catch macros work on device

d11755c

update

32bfdbf

fixes

7e3c597

simplify printing

6b3cfcd

fixes

3dab09d

davebayer added 3 commits May 14, 2026 20:35

more fixes

39c07c5

improve doc strings

1265a5c

fix clang tidy

9690649

davebayer force-pushed the c2h_test_macros branch from bd9a3ea to 896439c Compare May 14, 2026 18:37

Jacobfaib approved these changes May 14, 2026

View reviewed changes

alliepiper approved these changes May 14, 2026

View reviewed changes

bernhardmgruber approved these changes May 14, 2026

View reviewed changes

fixes

694ebaf

davebayer force-pushed the c2h_test_macros branch from 896439c to 694ebaf Compare May 14, 2026 20:35

fixes

490c531

davebayer enabled auto-merge (squash) May 14, 2026 20:40

coderabbitai Bot reviewed May 14, 2026

View reviewed changes

Comment thread c2h/include/c2h/catch2_test_macros.h

davebayer merged commit 16d667a into NVIDIA:main May 14, 2026
380 checks passed

This was referenced May 15, 2026

[libcu++] Remove CCCLRT_MEOW test macros #8996

Open

[cudax] Remove CUDAX_MEOW test macros #8998

Open

[cub] Replace assert with CHECK or REQUIRE in tests #8999

Open

[cub] Use REQUIRE_CUDART in catch2 tests #9000

Open

Conversation

davebayer commented May 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This comment has been minimized.

coderabbitai Bot commented May 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Walkthrough

Changes

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

This comment has been minimized.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Jacobfaib May 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

alliepiper left a comment

Choose a reason for hiding this comment

Uh oh!

bernhardmgruber left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions Bot commented May 14, 2026

🥳 CI Workflow Results

🟩 Finished in 2h 32m: Pass: 100%/358 | Total: 8d 09h | Max: 2h 31m | Hits: 68%/684690

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

davebayer commented May 12, 2026 •

edited

Loading

coderabbitai Bot commented May 14, 2026 •

edited

Loading

Jacobfaib May 14, 2026 •

edited

Loading