stillwater-sc / universal / 32610739250

85%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

feat(elreal): evidence-based precision defaults, and fix what the sweep exposed (#1382)

Resolves #1177 by running the characterization tool (#1040) and acting on what it
measured. The answer is not the one the issue anticipated.

THE PREMISE CHANGED. #1177 asks to read the tool's "saturation knob -- where accuracy
stops improving" and set defaults from it. elreal no longer has one. The caps that
existed when the issue was filed were all removed in v4.9.0-v4.9.3 (#1076, #1371,
#1373), and accuracy is now linear in the knob and unbounded -- measured on sqrt(2),
double host:

    depth    2    3    4    5    6    7    8
    digits  32   49   66   83   98  115  131

16.4 digits per block against a predicted k*log10(2) = 15.95, with no plateau at any
depth swept. So the choice of default is a policy question, not a measurement.

THE TOOL WAS REPORTING THE OPERATOR'S OWN SWEEP LIMIT AS A RESULT. It picked "first
knob within 95% of the best digits" as the saturation point; for a series still
climbing, the best IS the last row, so it always fired at maxDepth. Every elreal
function claimed to saturate at whatever depth you happened to sweep to, and so did
most of ereal's. It now says which it observed:

    saturates ~depth K (D digits)
    NO saturation through depth K (D digits, still climbing)

The knee had the same flaw -- digits/log10(time) rises monotonically when accuracy
never plateaus -- and is likewise only reported when a plateau was seen. Under the
corrected reporting only ereal's sqrt genuinely saturates (N=12, 319 digits); its
transcendentals are still climbing at N=16.

kElrealDefaultPrecision 8 -> 32, ~510 decimal digits on a double host. It governs the
class facade only -- an elreal's own _depth for boundary ops and facade arithmetic;
the free ZBCL math functions carry their own depth arguments. Cost is roughly linear
in the knob; on the facade at -O2, double host, 8 -> 32 blocks: a/b 0.114 -> 0.576 ms,
a*b 0.240 -> 1.004 ms, a+b 0.022 -> 0.0... (continued)

Coverage Stats

3 of 3 new or added lines in 1 file covered. (100.0%)

8 existing lines in 1 file now uncovered.

43129 of 50519 relevant lines covered (85.37%)

7189054.34 hits per line