dzmingli_vs_floating/bench API

Luna-Flow/diff_bench/dzmingli_vs_floating/bench is an executable package. It has no public items; its interface is the command line and the files it writes. It uses the library package dzmingli_vs_floating.

Command

moon run --release src/dzmingli_vs_floating/bench --target native

The executable takes no arguments. The replay commands recorded in the JSONL append operation, size and implementation arguments; the executable ignores them and reruns the whole experiment.

Experiments

It runs run_mare_benchmark six times with default_benchmark_protocol() and seed 0xDEC1A1, three datasets per size:

Run	Operations	Coefficient digits	Scopes
scaling	add, subtract, multiply, divide, compare	1, 4, 16, 64, 256, 1,024, 4,096, 8,192, 10,000	arithmetic_only, full_path
stress	add, subtract, divide, compare	16,384, 20,000	arithmetic_only, full_path
extended	the 14 other exact operations	1, 16, 64, 256, 1,024	arithmetic_only, full_path

Output

• Standard output: the Mare Mark JSONL of every experiment (validation, calibration, observation, summary records), each followed by one "comparison" record per PerformanceResult.
• artifacts/dzmingli_vs_floating/scaling.html: a self-contained HTML report written with mare_performance_report_html. The directories are created when missing.

Exit status

The executable aborts with decimal differential validation failed when any validation failed, after the report has been written. A complete run without failures exits normally.

Targets

main.mbt is compiled for native only. On js, wasm, wasm-gc and llvm the package builds main_unimplemented.mbt, which prints that the runner requires native and exits.