Luna Flow
English

floating_vs_decmial_x/bench_common tutorial

This page runs the bench_common executable of floating_vs_decmial_x, keeps its records, and checks that the run is complete before you read its numbers.

Quick start

From the repository root:

moon run --release src/floating_vs_decmial_x/bench_common --target native \
  > artifacts/floating_vs_decmial_x/common_digits.jsonl

The command compiles a release build for native, validates every dataset against the oracle, measures both implementations and writes artifacts/floating_vs_decmial_x/common_digits.html. Open the HTML file in a browser.

Everyday tasks

Record the host

Mare Mark records only the facts you give it. Set them on the command line:

MARE_CPU="Apple M4" MARE_OS="macOS 26.5" MARE_BUILD_MODE=release \
  moon run --release src/floating_vs_decmial_x/bench_common --target native

Unset variables are recorded as unknown or unspecified.

Check that the run is complete

Every Mare Mark summary record must contain "complete":true, and every validation record must have "status":"valid" (for DzmingLi, failures from 4,096 digits are expected in the scaling run):

grep -c '"type":"validation"' artifacts/floating_vs_decmial_x/common_digits.jsonl
grep '"type":"validation"' artifacts/floating_vs_decmial_x/common_digits.jsonl | grep -vc '"status":"valid"'

Read the comparison records

Each "comparison" line is one operation and size. The *_median_us fields are median microseconds per operation, the speedup field is the GDA median over the X median, and decision applies the 3 % threshold.

Going further

Render publication figures from the JSONL with the Python layouts in tools/, as shown in the package tutorial. To change sizes or operations, edit main.mbt; the library functions take the operation list and the expanded size list as arguments.

Common pitfalls

  • Run with --release and --target native. Debug builds and other targets do not produce comparable numbers, and other targets do not run at all.
  • Do not merge records from different targets or hosts into one report.
  • A nonzero exit after a complete report means a validation failed; it is not a crash of the measurement.

Next steps