SEP 46: Calibrating the native backend to its machine #579

Tracking issue for SEP 46 — Calibrating the native CPU backend to its machine (Draft).

The native backend's CPU settings are constants: about seventeen operation thresholds, the medium width, the compact team's cap, the waiting policy and FFTW's planning effort. PR #576 measured them on six machine families and found the best value of each to depend on the machine. For example, dividing the thresholds by 16 ran an Apple M5 Max's cases 13 % faster but slowed some of a Xeon Gold's by 1.4–1.7×. That is why PR #576 needed per-platform switches.

SEP 46 proposes:

• solprop calibrate: a bounded one-off measurement (--quick about 20 s and automatic on first use, standard about a minute, --thorough about five minutes for benchmarks) of waiting, team width, core kinds, operation crossovers and FFTW plans, validated against the built-in table before use. FFTW efforts are tried as a ladder under time limits and kept only when the plan wins inside a step (no threaded patient planning, and no wisdom from losing plans);
• a per-user store in the platform cache directory (SOLPROP_CACHE_DIR overrides it), keyed by a machine and build fingerprint and holding FFTW wisdom by default, with read-only directories for clusters;
• an optional refinement for one simulation, keyed by a problem signature and reused by scans. This covers SEP 4's use case, and the SEP proposes superseding SEP 4 (SEP 4: Bounded execution autotuning #115).
• GPU builds' run-time settings, fitted when a rented device is bootstrapped (phase 6).

Related: #328, #329, #332, #81, #333, #136, #229, #303, #560.