Crate launchbound_bench 

Measurement: bench plans (plan.v1), the PTX runner (CUDA driver API via dlopen — builds anywhere, times only where a driver exists), and interval statistics. Every timing carries warmup, repeats, median, a 95% CI, and the outlier rule; overlapping intervals are reported as indistinguishable, never ranked (docs/BENCHMARKING.md).

Re-exports

pub use plan::ArgSpec;

pub use plan::BenchPlan;

pub use plan::BenchSpec;

pub use plan::Candidate;

pub use plan::PlanError;

pub use run::CandidateResult;

pub use run::Results;

pub use run::run_plan;

pub use stats::Summary;

pub use stats::indistinguishable;

pub use stats::summarize;

Modules

cuda

Minimal CUDA driver API surface, loaded at runtime with dlopen so the crate builds (and its tests run) on machines with no CUDA at all — including CI and the Mac. Only the box can actually produce a timing.

plan

Bench plans (plan.v1): everything the runner needs, self-contained — candidate configs, PTX paths, launch geometry, and the argument layout matching cuda-oxide’s PTX parameter lowering (each slice becomes a ptr, len pair of .param slots; scalars are single slots).

run

Plan execution: resumable, checkpointed after every candidate (the box dies — idle guard, dead-man switch, spot reclaim — so the harness is resumable or it is broken, docs/ARCHITECTURE.md), with a CPU heartbeat so a GPU-bound sweep never looks idle to the 30-minute CPU alarm.

stats

Interval statistics (docs/BENCHMARKING.md): a benchmark that reports a mean and no interval is not evidence. Median with a distribution-free 95% CI (order statistics), Tukey-fence outlier rejection, and an overlap test — configurations whose intervals overlap are indistinguishable, never ranked.