A JavaScript benchmark that measures the accuracy and performance of Math functions across engines. Unlike most benchmarks that focus solely on speed, js-math-bench treats correctness as the primary metric.

22 Math functions: acos, acosh, asin, asinh, atan, atan2, atanh, cbrt, cos, cosh, exp, expm1, log, log1p, log2, log10, pow, sin, sinh, sqrt, tan, tanh

For each function, the benchmark tests ~1300 inputs:

• Worst-case inputs (~1000) from the core-math project -- inputs where the exact result falls closest to the midpoint between two representable doubles, making correct rounding hardest.
• Systematic inputs (~250) evenly spaced across each function's domain, plus values near interesting points (domain boundaries, +-1, pi, etc.)
• Edge cases (~30) including denormals, very small/large values, and values near domain boundaries. (NaN/Inf/signed-zero inputs are excluded since those are spec compliance, not numerical accuracy.)

Expected results are pre-computed using Python's mpmath library at 200 digits of precision, then correctly rounded to double. The benchmark compares each engine's output and reports ULP (Units in Last Place) error.

• Accuracy score (0-100): 100 / (1 + mean ULP). Every ULP of error counts: perfect = 100, mean ULP 0.1 = 90.9, mean ULP 0.5 = 66.7, mean ULP 1.0 = 50. Since this uses the mean (not median), a few large errors pull the score down hard.
• Performance bonus (0-20): 20 * min(1, ops_per_sec / 1000M). Rewards throughput but capped so speed can't compensate for inaccuracy.
• Per-function total: accuracy + performance bonus (max 120).
• Overall score: geometric mean across all functions.

# Serve the directory
python3 -m http.server 8000
# Open http://localhost:8000 -- benchmark runs automatically

node runner-cli.js

For other JS engines (d8, jsc, etc.), generate a standalone single-file bundle:

node bundle.js > bench-standalone.js
d8 bench-standalone.js
jsc bench-standalone.js

cd c && make && ./bench-libm

The pre-computed test_data.json is checked into the repo. To regenerate:

git clone --depth 1 https://gitlab.inria.fr/core-math/core-math.git
uvx --with mpmath python3 generate_test_data.py