perf_runner API

Luna-Flow/linear-algebra/perf_runner is an executable that runs one benchmark case and prints a JSON record: either a diagnostic checksum or a list of timing samples. It has no public library items.

Source: src/perf_runner/main.mbt. The measurement method is in the perf_runner design.

Public items

None; the package is pkgtype(kind: "executable").

Command line

moon run src/perf_runner --target native <case-id> [--case-file <path>] [--repeat <n>] [--samples <n>] [--warmup <n>]

Argument	Default	Meaning
<case-id>	required	a case id registered in perf_support
--case-file <path>	bench/datasets/cases/<case-id>.json	fixture file; regenerated if missing
--repeat <n>	1	calls per timed sample, $n > 0$
--samples <n>	0	number of timed samples; 0 selects diagnostic mode
--warmup <n>	0	untimed warm-up rounds before sampling

Invalid or missing arguments abort with a usage message.

Output

Diagnostic mode (--samples 0) runs the case repeat times and prints case_diagnostic_payload: {"kind":"diagnostic", ..., "checksum":"<u64>"}. The checksum of repeat runs is the fold of the individual checksums.

Sampling mode prints {"kind":"benchmark", ..., "repeat":<n>, "samples":[t1, t2, ...]} where each $t_i$ is the elapsed wall-clock time of one sample divided by repeat, in nanoseconds.

For cases with mutation_policy = "scratch_per_sample", the inputs are copied for each call before the timer starts.