nirs4all-methods · cross-binding benchmark

✓ exact (≤ 1e-6) ≈ drift (≤ 1e-3) ✗ divergent (> 1e-3) ! error … deferred for catch-up run NR not run in this snapshot - not available in lib ⇄ cross-check (secondary oracle or documented noncanonical/RNG/model convention; informational) δ 1e-15 δ = relative RMSE (lower is better, ≈1e-12 = exact machine parity) ∩ J 0.75 J = Jaccard feature-set overlap for variable selectors (higher is better, 1.00 = identical selection) BD J 0.40 by design: different selector RNG/noise/model semantics; exact mask match is not expected 42.5 fastest row row best · language best · ≤15% · ≤50% · ≤100% · rest cell X.X ✓✓: gate 1 (binding ≡ cpp) · gate 2 (cpp ≈ external ref) click header to sort · ⇧+click for descending

— —

methodology

Reference: pls4all.cpp native column (resolved from data)
Parity tolerance: 1e-6 exact, 1e-3 drift threshold
Every backend reads the same orchestrator-generated CSV dataset
Adaptive timing: run #1 is warmstart; 2 runs report run #2, 3 runs average runs #2-#3, 10/20/40 runs report the median after warmstart
Per-cell timeout guard: 24h
Thread control via OMP_NUM_THREADS · OPENBLAS_NUM_THREADS · MKL_NUM_THREADS · BLIS_NUM_THREADS + binding-specific thread controls

See the full methodology page for reference policy and per-algo overrides.

host

parity verdict

Each cell shows <duration_ms> <icon> or <speed-up×> <icon> vs the native C++ column.

✓ exact — ≤ 1e-6 vs reference
≈ drift — ≤ 1e-3 and not exact
✗ divergent — > 1e-3
! error — dependency, runtime, crash or exception
… deferred — intentionally skipped from the CRAN-fast benchmark; catch-up runs fill it later
- not available in lib — the library does not expose that algorithm

Violet = fastest row cell. Blue = fastest within language; green/orange/red compare each cell time to that language winner.

methodology

host

parity verdict

Backends & libraries