[STF] Low level interface for the cuda_kernel(_chain) construct by caugonnet · Pull Request #5319 · NVIDIA/cccl

caugonnet · 2025-07-20T06:43:31Z

Description

Implement a low-level interface for the cuda_kernel(_chain) construct so that it can later be called from C/python

closes

Checklist

New or existing tests cover these changes.
The documentation is up to date with these changes.

…rs used by kernels

copy-pr-bot · 2025-07-20T06:43:35Z

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.

caugonnet · 2025-07-23T16:26:39Z

/ok to test 316167f

github-actions · 2025-07-23T17:01:37Z

🟩 CI finished in 32m 04s: Pass: 100%/32 | Total: 7h 03m | Avg: 13m 13s | Max: 29m 27s | Hits: 75%/15978

🟩 cudax: Pass: 100%/28 | Total: 6h 48m | Avg: 14m 35s | Max: 29m 27s | Hits: 75%/15978

🟩 cpu
  🟩 amd64              Pass: 100%/24  | Total:  5h 52m | Avg: 14m 41s | Max: 29m 27s | Hits:  76%/13522 
  🟩 arm64              Pass: 100%/4   | Total: 56m 00s | Avg: 14m 00s | Max: 15m 47s | Hits:  69%/2456  
🟩 ctk
  🟩 12.0               Pass: 100%/3   | Total: 37m 39s | Avg: 12m 33s | Max: 14m 38s | Hits:  75%/1537  
  🟩 12.9               Pass: 100%/25  | Total:  6h 10m | Avg: 14m 49s | Max: 29m 27s | Hits:  75%/14441 
🟩 cudacxx
  🟩 nvcc12.0           Pass: 100%/3   | Total: 37m 39s | Avg: 12m 33s | Max: 14m 38s | Hits:  75%/1537  
  🟩 nvcc12.9           Pass: 100%/25  | Total:  6h 10m | Avg: 14m 49s | Max: 29m 27s | Hits:  75%/14441 
🟩 cudacxx_family
  🟩 nvcc               Pass: 100%/28  | Total:  6h 48m | Avg: 14m 35s | Max: 29m 27s | Hits:  75%/15978 
🟩 cxx
  🟩 Clang14            Pass: 100%/2   | Total: 24m 59s | Avg: 12m 29s | Max: 13m 21s | Hits:  70%/1230  
  🟩 Clang15            Pass: 100%/1   | Total: 15m 11s | Avg: 15m 11s | Max: 15m 11s | Hits:  70%/614   
  🟩 Clang16            Pass: 100%/1   | Total: 15m 15s | Avg: 15m 15s | Max: 15m 15s | Hits:  70%/614   
  🟩 Clang17            Pass: 100%/1   | Total: 15m 21s | Avg: 15m 21s | Max: 15m 21s | Hits:  70%/614   
  🟩 Clang18            Pass: 100%/1   | Total: 15m 02s | Avg: 15m 02s | Max: 15m 02s | Hits:  70%/614   
  🟩 Clang19            Pass: 100%/4   | Total: 49m 06s | Avg: 12m 16s | Max: 15m 12s | Hits:  77%/2456  
  🟩 GCC10              Pass: 100%/2   | Total: 30m 31s | Avg: 15m 15s | Max: 15m 53s | Hits:  69%/1230  
  🟩 GCC11              Pass: 100%/1   | Total: 16m 05s | Avg: 16m 05s | Max: 16m 05s | Hits:  69%/614   
  🟩 GCC12              Pass: 100%/1   | Total: 17m 47s | Avg: 17m 47s | Max: 17m 47s | Hits:  69%/614   
  🟩 GCC13              Pass: 100%/8   | Total:  1h 47m | Avg: 13m 23s | Max: 16m 58s | Hits:  77%/4912  
  🟩 MSVC14.39          Pass: 100%/1   | Total: 11m 23s | Avg: 11m 23s | Max: 11m 23s | Hits:  95%/309   
  🟩 MSVC14.43          Pass: 100%/3   | Total: 33m 06s | Avg: 11m 02s | Max: 11m 47s | Hits:  95%/933   
  🟩 NVHPC25.5          Pass: 100%/2   | Total: 57m 33s | Avg: 28m 46s | Max: 29m 27s | Hits:  67%/1224  
🟩 cxx_family
  🟩 Clang              Pass: 100%/10  | Total:  2h 14m | Avg: 13m 29s | Max: 15m 21s | Hits:  73%/6142  
  🟩 GCC                Pass: 100%/12  | Total:  2h 51m | Avg: 14m 17s | Max: 17m 47s | Hits:  74%/7370  
  🟩 MSVC               Pass: 100%/4   | Total: 44m 29s | Avg: 11m 07s | Max: 11m 47s | Hits:  95%/1242  
  🟩 NVHPC              Pass: 100%/2   | Total: 57m 33s | Avg: 28m 46s | Max: 29m 27s | Hits:  67%/1224  
🟩 gpu
  🟩 h100               Pass: 100%/2   | Total: 19m 51s | Avg:  9m 55s | Max: 12m 33s | Hits:  84%/1228  
  🟩 rtx2080            Pass: 100%/26  | Total:  6h 28m | Avg: 14m 56s | Max: 29m 27s | Hits:  74%/14750 
🟩 jobs
  🟩 Build              Pass: 100%/25  | Total:  6h 22m | Avg: 15m 18s | Max: 29m 27s | Hits:  71%/14136 
  🟩 Test               Pass: 100%/3   | Total: 25m 54s | Avg:  8m 38s | Max: 10m 34s | Hits:  99%/1842  
🟩 sm
  🟩 90                 Pass: 100%/2   | Total: 19m 51s | Avg:  9m 55s | Max: 12m 33s | Hits:  84%/1228  
  🟩 90;90a             Pass: 100%/2   | Total: 25m 28s | Avg: 12m 44s | Max: 14m 39s | Hits:  78%/925   
  🟩 100;120            Pass: 100%/2   | Total: 25m 25s | Avg: 12m 42s | Max: 14m 55s | Hits:  78%/925   
🟩 std
  🟩 17                 Pass: 100%/3   | Total: 54m 58s | Avg: 18m 19s | Max: 28m 06s | Hits:  69%/1840  
  🟩 20                 Pass: 100%/25  | Total:  5h 53m | Avg: 14m 08s | Max: 29m 27s | Hits:  75%/14138

🟩 packaging: Pass: 100%/4 | Total: 14m 39s | Avg: 3m 39s | Max: 3m 57s

🟩 cpu
  🟩 amd64              Pass: 100%/4   | Total: 14m 39s | Avg:  3m 39s | Max:  3m 57s
🟩 ctk
  🟩 12.0               Pass: 100%/2   | Total:  7m 22s | Avg:  3m 41s | Max:  3m 57s
  🟩 12.9               Pass: 100%/2   | Total:  7m 17s | Avg:  3m 38s | Max:  3m 55s
🟩 cudacxx
  🟩 nvcc12.0           Pass: 100%/2   | Total:  7m 22s | Avg:  3m 41s | Max:  3m 57s
  🟩 nvcc12.9           Pass: 100%/2   | Total:  7m 17s | Avg:  3m 38s | Max:  3m 55s
🟩 cudacxx_family
  🟩 nvcc               Pass: 100%/4   | Total: 14m 39s | Avg:  3m 39s | Max:  3m 57s
🟩 cxx
  🟩 Clang14            Pass: 100%/1   | Total:  3m 57s | Avg:  3m 57s | Max:  3m 57s
  🟩 Clang19            Pass: 100%/1   | Total:  3m 22s | Avg:  3m 22s | Max:  3m 22s
  🟩 GCC12              Pass: 100%/1   | Total:  3m 25s | Avg:  3m 25s | Max:  3m 25s
  🟩 GCC13              Pass: 100%/1   | Total:  3m 55s | Avg:  3m 55s | Max:  3m 55s
🟩 cxx_family
  🟩 Clang              Pass: 100%/2   | Total:  7m 19s | Avg:  3m 39s | Max:  3m 57s
  🟩 GCC                Pass: 100%/2   | Total:  7m 20s | Avg:  3m 40s | Max:  3m 55s
🟩 gpu
  🟩 rtx2080            Pass: 100%/4   | Total: 14m 39s | Avg:  3m 39s | Max:  3m 57s
🟩 jobs
  🟩 Test               Pass: 100%/4   | Total: 14m 39s | Avg:  3m 39s | Max:  3m 57s

👃 Inspect Changes

Modifications in project?

	Project
	CCCL Infrastructure
	CCCL Packaging
	libcu++
	CUB
	Thrust
+/-	CUDA Experimental
	stdpar
	python
	CCCL C Parallel Library
	Catch2Helper

Modifications in project or dependencies?

	Project
	CCCL Infrastructure
+/-	CCCL Packaging
	libcu++
	CUB
	Thrust
+/-	CUDA Experimental
	stdpar
	python
	CCCL C Parallel Library
	Catch2Helper

🏃‍ Runner counts (total jobs: 32)

#	Runner
17	`linux-amd64-cpu16`
6	`linux-amd64-gpu-rtx2080-latest-1`
4	`linux-arm64-cpu16`
4	`windows-amd64-cpu16`
1	`linux-amd64-gpu-h100-latest-1`

cudax/include/cuda/experimental/__stf/internal/cuda_kernel_scope.cuh

andralex · 2025-07-24T10:11:53Z

cudax/include/cuda/experimental/__stf/internal/cuda_kernel_scope.cuh

  dim3 gridDim;
  dim3 blockDim;


Do these have a default ctor, and if not, should we add a = {}?

cudax/include/cuda/experimental/__stf/internal/cuda_kernel_scope.cuh

andralex · 2025-07-24T10:26:43Z

Made a pass, good stuff @caugonnet. Thanks @bernhardmgruber for taking a look!

…vebayer

…vebayer for finding this)

caugonnet · 2025-08-04T14:12:05Z

/ok to test c3a4c99

copy-pr-bot · 2025-08-04T14:12:08Z

/ok to test c3a4c99

@caugonnet, there was an error processing your request: E2

See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/2/

…pe.cuh Co-authored-by: Andrei Alexandrescu <andrei@erdani.com>

caugonnet · 2025-08-04T14:16:08Z

/ok to test 9a9cde3

copy-pr-bot · 2025-08-04T14:16:11Z

/ok to test 9a9cde3

@caugonnet, there was an error processing your request: E2

See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/2/

caugonnet · 2025-08-04T14:20:05Z

/ok to test dec6b46

github-actions · 2025-08-04T14:54:50Z

🟩 CI finished in 32m 30s: Pass: 100%/32 | Total: 7h 41m | Avg: 14m 24s | Max: 32m 25s | Hits: 72%/15390

🟩 cudax: Pass: 100%/28 | Total: 7h 27m | Avg: 15m 59s | Max: 32m 25s | Hits: 72%/15390

🟩 cpu
  🟩 amd64              Pass: 100%/24  | Total:  6h 24m | Avg: 16m 01s | Max: 32m 25s | Hits:  72%/13018 
  🟩 arm64              Pass: 100%/4   | Total:  1h 02m | Avg: 15m 42s | Max: 16m 20s | Hits:  68%/2372  
🟩 ctk
  🟩 12.0               Pass: 100%/3   | Total: 37m 52s | Avg: 12m 37s | Max: 14m 10s | Hits:  72%/1474  
  🟩 12.9               Pass: 100%/25  | Total:  6h 49m | Avg: 16m 23s | Max: 32m 25s | Hits:  72%/13916 
🟩 cudacxx
  🟩 nvcc12.0           Pass: 100%/3   | Total: 37m 52s | Avg: 12m 37s | Max: 14m 10s | Hits:  72%/1474  
  🟩 nvcc12.9           Pass: 100%/25  | Total:  6h 49m | Avg: 16m 23s | Max: 32m 25s | Hits:  72%/13916 
🟩 cudacxx_family
  🟩 nvcc               Pass: 100%/28  | Total:  7h 27m | Avg: 15m 59s | Max: 32m 25s | Hits:  72%/15390 
🟩 cxx
  🟩 Clang14            Pass: 100%/2   | Total: 28m 44s | Avg: 14m 22s | Max: 16m 29s | Hits:  69%/1188  
  🟩 Clang15            Pass: 100%/1   | Total: 17m 26s | Avg: 17m 26s | Max: 17m 26s | Hits:  67%/593   
  🟩 Clang16            Pass: 100%/1   | Total: 17m 03s | Avg: 17m 03s | Max: 17m 03s | Hits:  61%/593   
  🟩 Clang17            Pass: 100%/1   | Total: 16m 01s | Avg: 16m 01s | Max: 16m 01s | Hits:  66%/593   
  🟩 Clang18            Pass: 100%/1   | Total: 17m 03s | Avg: 17m 03s | Max: 17m 03s | Hits:  65%/593   
  🟩 Clang19            Pass: 100%/4   | Total: 54m 48s | Avg: 13m 42s | Max: 15m 38s | Hits:  76%/2372  
  🟩 GCC10              Pass: 100%/2   | Total: 31m 56s | Avg: 15m 58s | Max: 17m 46s | Hits:  68%/1188  
  🟩 GCC11              Pass: 100%/1   | Total: 17m 31s | Avg: 17m 31s | Max: 17m 31s | Hits:  61%/593   
  🟩 GCC12              Pass: 100%/1   | Total: 19m 32s | Avg: 19m 32s | Max: 19m 32s | Hits:  68%/593   
  🟩 GCC13              Pass: 100%/8   | Total:  1h 56m | Avg: 14m 34s | Max: 18m 36s | Hits:  75%/4744  
  🟩 MSVC14.39          Pass: 100%/1   | Total: 11m 27s | Avg: 11m 27s | Max: 11m 27s | Hits:  89%/288   
  🟩 MSVC14.43          Pass: 100%/3   | Total: 38m 02s | Avg: 12m 40s | Max: 13m 22s | Hits:  91%/870   
  🟩 NVHPC25.5          Pass: 100%/2   | Total:  1h 01m | Avg: 30m 42s | Max: 32m 25s | Hits:  61%/1182  
🟩 cxx_family
  🟩 Clang              Pass: 100%/10  | Total:  2h 31m | Avg: 15m 06s | Max: 17m 26s | Hits:  70%/5932  
  🟩 GCC                Pass: 100%/12  | Total:  3h 05m | Avg: 15m 27s | Max: 19m 32s | Hits:  72%/7118  
  🟩 MSVC               Pass: 100%/4   | Total: 49m 29s | Avg: 12m 22s | Max: 13m 22s | Hits:  90%/1158  
  🟩 NVHPC              Pass: 100%/2   | Total:  1h 01m | Avg: 30m 42s | Max: 32m 25s | Hits:  61%/1182  
🟩 gpu
  🟩 h100               Pass: 100%/2   | Total: 21m 15s | Avg: 10m 37s | Max: 12m 13s | Hits:  84%/1186  
  🟩 rtx2080            Pass: 100%/26  | Total:  7h 06m | Avg: 16m 23s | Max: 32m 25s | Hits:  71%/14204 
🟩 jobs
  🟩 Build              Pass: 100%/25  | Total:  6h 56m | Avg: 16m 40s | Max: 32m 25s | Hits:  68%/13611 
  🟩 Test               Pass: 100%/3   | Total: 30m 34s | Avg: 10m 11s | Max: 12m 37s | Hits:  99%/1779  
🟩 sm
  🟩 90                 Pass: 100%/2   | Total: 21m 15s | Avg: 10m 37s | Max: 12m 13s | Hits:  84%/1186  
  🟩 90;90a             Pass: 100%/2   | Total: 29m 09s | Avg: 14m 34s | Max: 15m 47s | Hits:  76%/883   
  🟩 100;120            Pass: 100%/2   | Total: 27m 14s | Avg: 13m 37s | Max: 15m 46s | Hits:  74%/883   
🟩 std
  🟩 17                 Pass: 100%/3   | Total:  1h 00m | Avg: 20m 02s | Max: 28m 59s | Hits:  65%/1777  
  🟩 20                 Pass: 100%/25  | Total:  6h 27m | Avg: 15m 29s | Max: 32m 25s | Hits:  73%/13613

🟩 packaging: Pass: 100%/4 | Total: 13m 46s | Avg: 3m 26s | Max: 3m 31s

🟩 cpu
  🟩 amd64              Pass: 100%/4   | Total: 13m 46s | Avg:  3m 26s | Max:  3m 31s
🟩 ctk
  🟩 12.0               Pass: 100%/2   | Total:  6m 52s | Avg:  3m 26s | Max:  3m 27s
  🟩 12.9               Pass: 100%/2   | Total:  6m 54s | Avg:  3m 27s | Max:  3m 31s
🟩 cudacxx
  🟩 nvcc12.0           Pass: 100%/2   | Total:  6m 52s | Avg:  3m 26s | Max:  3m 27s
  🟩 nvcc12.9           Pass: 100%/2   | Total:  6m 54s | Avg:  3m 27s | Max:  3m 31s
🟩 cudacxx_family
  🟩 nvcc               Pass: 100%/4   | Total: 13m 46s | Avg:  3m 26s | Max:  3m 31s
🟩 cxx
  🟩 Clang14            Pass: 100%/1   | Total:  3m 27s | Avg:  3m 27s | Max:  3m 27s
  🟩 Clang19            Pass: 100%/1   | Total:  3m 31s | Avg:  3m 31s | Max:  3m 31s
  🟩 GCC12              Pass: 100%/1   | Total:  3m 25s | Avg:  3m 25s | Max:  3m 25s
  🟩 GCC13              Pass: 100%/1   | Total:  3m 23s | Avg:  3m 23s | Max:  3m 23s
🟩 cxx_family
  🟩 Clang              Pass: 100%/2   | Total:  6m 58s | Avg:  3m 29s | Max:  3m 31s
  🟩 GCC                Pass: 100%/2   | Total:  6m 48s | Avg:  3m 24s | Max:  3m 25s
🟩 gpu
  🟩 rtx2080            Pass: 100%/4   | Total: 13m 46s | Avg:  3m 26s | Max:  3m 31s
🟩 jobs
  🟩 Test               Pass: 100%/4   | Total: 13m 46s | Avg:  3m 26s | Max:  3m 31s

👃 Inspect Changes

Modifications in project?

	Project
	CCCL Infrastructure
	CCCL Packaging
	libcu++
	CUB
	Thrust
+/-	CUDA Experimental
	stdpar
	python
	CCCL C Parallel Library
	Catch2Helper

Modifications in project or dependencies?

	Project
	CCCL Infrastructure
+/-	CCCL Packaging
	libcu++
	CUB
	Thrust
+/-	CUDA Experimental
	stdpar
	python
	CCCL C Parallel Library
	Catch2Helper

🏃‍ Runner counts (total jobs: 32)

#	Runner
17	`linux-amd64-cpu16`
6	`linux-amd64-gpu-rtx2080-latest-1`
4	`linux-arm64-cpu16`
4	`windows-amd64-cpu16`
1	`linux-amd64-gpu-h100-latest-1`

@davebayer

…IA#5319) * Allow CUfunction (driver API) in the cuda_kernel(_chain) API * clang-format * We have a std::tuple not a cuda::std::tuple (yet) * If CUDASTF_CUDA_KERNEL_DEBUG is set, we display the number of registers used by kernels * Support CUkernel in addition to CUfunction * Add a test with CUfunction and CUkernel * Check whether CUkernel is supported * use _CCCL_ASSERT instead of assert to avoid an unused variable error * cudaGetKernel was added in CUDA 12.1 * clang-format * Extract the start and end phase of the ->* operator * There is no need to store untyped_t as we now store the task with its type * Implement the low level interface for cuda_kernel(_chain) with a way to avoid using the ->* operator * - Add a test to ensure we can put no arguments in the cuda_kernel_desc constructor - Implement a low level API to describe cuda_kernel_desc with an array of pointers rather than a variadic interface (and use it in a test) * simpler code, and do not check for CUDA_VERSION >= 12 * Simpler code Co-authored-by: Andrei Alexandrescu <andrei@erdani.com> * Update cudax/include/cuda/experimental/__stf/internal/cuda_kernel_scope.cuh Co-authored-by: Andrei Alexandrescu <andrei@erdani.com> * clang-format * Do not test for CUDA_VERSION >= 12 * Add missing const * add missing template * Use _CCCL_CTK_AT_LEAST * replace std::visit by std::get_if in get_num_registers * use ::std::get_if instead of ::std::visit * Add some tests for get_num_registers which actually fails (thanks @davebayer for finding this) * Fix the method to get the number of registers for CUkernel * Update cudax/include/cuda/experimental/__stf/internal/cuda_kernel_scope.cuh Co-authored-by: Andrei Alexandrescu <andrei@erdani.com> --------- Co-authored-by: Andrei Alexandrescu <andrei@erdani.com>

caugonnet and others added 19 commits July 11, 2025 13:54

Allow CUfunction (driver API) in the cuda_kernel(_chain) API

f2a3777

clang-format

b3304a1

We have a std::tuple not a cuda::std::tuple (yet)

8651e9f

Merge branch 'main' into stf_cufunction_cuda_kernel

0a0cd17

Merge branch 'main' into stf_cufunction_cuda_kernel

f9bf270

If CUDASTF_CUDA_KERNEL_DEBUG is set, we display the number of registe…

fd42c70

…rs used by kernels

Merge branch 'main' into stf_cufunction_cuda_kernel

e12de1c

Merge branch 'main' into stf_cufunction_cuda_kernel

a8aa5f1

Merge branch 'main' into stf_cufunction_cuda_kernel

74a4577

Merge branch 'main' into stf_cufunction_cuda_kernel

b809199

Support CUkernel in addition to CUfunction

f36fcd0

Add a test with CUfunction and CUkernel

b2002d8

Merge branch 'main' into stf_cufunction_cuda_kernel

6db15e2

Check whether CUkernel is supported

22c758a

use _CCCL_ASSERT instead of assert to avoid an unused variable error

02ded8b

cudaGetKernel was added in CUDA 12.1

ebeb703

clang-format

2f298b7

Merge branch 'main' into stf_cufunction_cuda_kernel

af8eec6

Extract the start and end phase of the ->* operator

3d9b7a5

caugonnet requested review from a team as code owners July 20, 2025 06:43

github-project-automation bot added this to CCCL Jul 20, 2025

github-project-automation bot moved this to Todo in CCCL Jul 20, 2025

caugonnet requested review from gonidelis and wmaxey July 20, 2025 06:43

caugonnet requested a review from griwes July 20, 2025 06:43

caugonnet self-assigned this Jul 20, 2025

cccl-authenticator-app bot moved this from Todo to In Review in CCCL Jul 20, 2025

caugonnet added 5 commits July 23, 2025 16:14

add missing template

01d638f

Use _CCCL_CTK_AT_LEAST

d79090b

replace std::visit by std::get_if in get_num_registers

e0672ca

use ::std::get_if instead of ::std::visit

4c8cfaf

Merge branch 'stf_cufunction_cuda_kernel' into stf_cuda_kernel_lowlevel

316167f