GPU: skip events for default kernel launches

[joseph-isaacs]

Independent of the other follow-up PRs; based directly on #9147. This draft contains one reviewable GPU correctness or performance concern.

What

• Make kernel event recording conditional on the launch strategy.
• Record no before/after events for normal execution.
• Preserve timed events for tracing and profiling strategies.

Why this is needed

The default launch strategy records before/after CUDA events around every kernel even though its completion hook is a no-op. Those events therefore cannot affect correctness or produce a metric. The profile observed 1,865 event records across the benchmark. This change skips them only for the default strategy and preserves event timing for tracing/profiling strategies, removing two unnecessary driver operations per normal launch.

Validation

Checked independently against this PR's current base:

• cargo check -p vortex-cuda --all-features

The combined implementation also passed Rust/CUDA formatting, 164 focused dynamic-dispatch tests, and 28 focused constant-array tests.

[codspeed-hq]

Merging this PR will degrade performance by 11.93%

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

❌ 2 regressed benchmarks
✅ 2041 untouched benchmarks
⏩ 46 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	words_gather_scalar[65536]	8.2 µs	9.4 µs	-12.74%
❌	WallTime	words_gather_dispatch[1024]	8 ns	9 ns	-11.11%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing joe/gpu-default-launch-no-events (79352b6) with develop (b825c4f)}

Footnotes

1. 46 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩