cudax/stf: cuda_try migration — stream_task (PR9) by andralex · Pull Request #10003 · NVIDIA/cccl

andralex · 2026-07-18T03:12:02Z

Summary

migrate throwing-capable CUDA calls in stream_task.cuh from cuda_safe_call to cuda_try
make timing-event setup and destruction exception-safe across stream tasks, host launches, and kernel launches
separate successful timing collection from failure cleanup and remove redundant timing_active state
use the capture-aware stream-device helper and scoped device activation for multi-place stream dependencies
simplify related timing predicates in graph tasks

Test plan

pre-commit run --files cudax/include/cuda/experimental/__stf/graph/graph_task.cuh cudax/include/cuda/experimental/__stf/internal/host_launch_scope.cuh cudax/include/cuda/experimental/__stf/internal/launch.cuh cudax/include/cuda/experimental/__stf/stream/stream_task.cuh
ninja -j 12 -C build/cudax-cpp20 cudax.test.stf.dot.with_events cudax.test.stf.cpp.task_get_stream cudax.test.stf.interface.host_launch_deps cudax.test.stf.stress.launch_overhead cudax.test.stf.places.execution_policy_kernel_launch_test
ctest --test-dir build/cudax-cpp20 -j 12 -R 'cudax\.test\.stf\.(examples\.01-axpy-places|places\.(cuda_stream_place|non_current_device))' --output-on-failure
CUDASTF_DOT_TIMING=1 ./build/cudax-cpp20/bin/cudax.test.stf.examples.01-axpy-places

Migrate throwing-capable stream task setup to cuda_try while retaining cuda_safe_call in noexcept scope guards. Make timing-event ownership exception-safe and use the capture-aware stream-device helper so partial failures restore device state and release events.

copy-pr-bot · 2026-07-18T03:12:05Z

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.

andralex · 2026-07-18T03:12:15Z

/ok to test caf42e7

coderabbitai · 2026-07-18T03:14:01Z

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: f8895329-5eb5-4b39-9cb5-e30b3c22e25a

📥 Commits

Reviewing files that changed from the base of the PR and between f89181b and 7e45b83.

📒 Files selected for processing (1)

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

🚧 Files skipped from review as they are similar to previous changes (1)

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

📝 Walkthrough

Summary by CodeRabbit

Bug Fixes
- Improved CUDA task timing reliability by creating timing events only when needed and ensuring they’re properly destroyed on both success and failure.
- Refined stream and graph task timing so elapsed-time reporting happens only when valid timing data is available.
- Strengthened synchronization and dependency insertion during launches, including safer handling of cooperative-group arrived lists and improved cleanup of temporary managed resources in error scenarios.

Walkthrough

Changes

STF timing paths now use conditionally created CUDA events with RAII cleanup and explicit success/failure handling. Stream dependency insertion is device-scoped, launch cleanup covers temporary and cooperative-group allocations, and graph task timing decisions combine scheduling with calibration state.

STF timing and cleanup

Layer / File(s)	Summary
Stream task event lifecycle `cudax/include/cuda/experimental/__stf/stream/stream_task.cuh`	Both stream task variants guard timing event creation, measurement, failure cleanup, and destruction; dependency insertion uses a device-scoped synchronization event.
Launch cleanup and timing flow `cudax/include/cuda/experimental/__stf/internal/launch.cuh`, `cudax/include/cuda/experimental/__stf/internal/host_launch_scope.cuh`	Launch paths separate success and failure cleanup, free temporary and arrived-list allocations, correct constructor and allocation sizing, and measure timing only when events exist.
Graph task timing decision `cudax/include/cuda/experimental/__stf/graph/graph_task.cuh`	Typed and untyped graph tasks compute `record_time` from scheduling and calibration state.

Possibly related PRs

NVIDIA/cccl#9303: Updates related STF event dependency insertion and guarded CUDA event lifecycles.

Suggested labels: stf

Suggested reviewers: caugonnet, srinivasyadav18, jacobfaib

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

Destroy timing events reliably and separate successful timing collection from failure cleanup across stream and launch tasks.

andralex · 2026-07-18T03:49:30Z

/ok to test 19bb001

coderabbitai

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

cudax/include/cuda/experimental/__stf/internal/launch.cuh (1)

399-403: 🩺 Stability & Availability | 🟠 Major | ⚡ Quick win

important: pass grid_size - 1 to deallocateManagedMemory here deallocateManagedMemory keys the pool by sz, so freeing the allocateManagedMemory(grid_size - 1) buffer with grid_size reinserts it under the wrong size and can later return a short buffer for a grid_size allocation.

🧹 Nitpick comments (1)

cudax/include/cuda/experimental/__stf/internal/launch.cuh (1)

389-389: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value

suggestion: Declare free_launch_temps as const; it is never reassigned. As per coding guidelines, “All variables that are not modified must be declared const.”

Source: Coding guidelines

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 56b6c5a9-8b62-4e1d-8591-f6c5a43c262b

📥 Commits

Reviewing files that changed from the base of the PR and between caf42e7 and 19bb001.

📒 Files selected for processing (4)

cudax/include/cuda/experimental/__stf/graph/graph_task.cuh
cudax/include/cuda/experimental/__stf/internal/host_launch_scope.cuh
cudax/include/cuda/experimental/__stf/internal/launch.cuh
cudax/include/cuda/experimental/__stf/stream/stream_task.cuh

Build the execution policy before starting the task so teardown guards cover occupancy failures, adopt arrived-list ownership atomically, and install the device-temp free guard before set_device_tmp.

The 3-arg constructor forwarded arg and streams in the wrong order to the primary constructor.

andralex · 2026-07-18T04:32:53Z

/ok to test f89181b

A SCOPE(fail) guard only runs during unwinding, so it can directly free an unadopted arrived list without tracking redundant state.

…-stream-task

andralex · 2026-07-18T04:59:04Z

/ok to test 7e45b83

github-actions · 2026-07-18T06:26:13Z

😬 CI Workflow Results

🟥 Finished in 1h 24m: Pass: 28%/57 | Total: 1d 11h | Max: 1h 24m | Hits: 14%/30082

See results here.

andralex requested a review from a team as a code owner July 18, 2026 03:12

andralex requested a review from caugonnet July 18, 2026 03:12

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

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

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

cudax/stf: make task timing cleanup exception-safe

19bb001

Destroy timing events reliably and separate successful timing collection from failure cleanup across stream and launch tasks.

coderabbitai Bot reviewed Jul 18, 2026

View reviewed changes

andralex added 2 commits July 18, 2026 00:32

cudax/stf: close launch exception-safety gaps

7a17bd9

Build the execution policy before starting the task so teardown guards cover occupancy failures, adopt arrived-list ownership atomically, and install the device-temp free guard before set_device_tmp.

cudax/stf: fix swapped streams/arg in launch delegating ctor

f89181b

The 3-arg constructor forwarded arg and streams in the wrong order to the primary constructor.

andralex added 2 commits July 18, 2026 00:57

cudax/stf: simplify launch failure guards

01833c5

A SCOPE(fail) guard only runs during unwinding, so it can directly free an unadopted arrived list without tracking redundant state.

Merge remote-tracking branch 'origin/main' into andralex/stf-cuda-try…

7e45b83

…-stream-task

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

cudax/stf: cuda_try migration — stream_task (PR9)#10003

cudax/stf: cuda_try migration — stream_task (PR9)#10003
andralex wants to merge 6 commits into
NVIDIA:mainfrom
andralex:andralex/stf-cuda-try-stream-task

andralex commented Jul 18, 2026 •

edited

Loading

Uh oh!

copy-pr-bot Bot commented Jul 18, 2026

Uh oh!

andralex commented Jul 18, 2026

Uh oh!

coderabbitai Bot commented Jul 18, 2026 •

edited

Loading

Summary by CodeRabbit

Walkthrough

Changes

Uh oh!

andralex commented Jul 18, 2026

Uh oh!

coderabbitai Bot left a comment

Uh oh!

andralex commented Jul 18, 2026

Uh oh!

andralex commented Jul 18, 2026

Uh oh!

github-actions Bot commented Jul 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

andralex commented Jul 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Uh oh!

copy-pr-bot Bot commented Jul 18, 2026

Uh oh!

andralex commented Jul 18, 2026

Uh oh!

coderabbitai Bot commented Jul 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Walkthrough

Changes

Uh oh!

andralex commented Jul 18, 2026

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

andralex commented Jul 18, 2026

Uh oh!

andralex commented Jul 18, 2026

Uh oh!

github-actions Bot commented Jul 18, 2026

😬 CI Workflow Results

🟥 Finished in 1h 24m: Pass: 28%/57 | Total: 1d 11h | Max: 1h 24m | Hits: 14%/30082

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

andralex commented Jul 18, 2026 •

edited

Loading

coderabbitai Bot commented Jul 18, 2026 •

edited

Loading