• Home
  • Features
  • Pricing
  • Docs
  • Announcements
  • Sign In

stillwater-sc / universal / 39202799770
89%
master: 84%

Build:
Build:
LAST BUILD BRANCH: main
DEFAULT BRANCH: master
Ran 19 Oct 2026 05:25AM UTC
Jobs 1
Files 754
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst

19 Oct 2026 05:06AM UTC coverage: 87.435% (-0.01%) from 87.447%
39202799770

push

github

web-flow
perf(efloat): log by a table and an atanh series, 6-10x faster; exp and log within 2x of mpmath (#1754)

log ran Newton on exp(y) - x = 0, about 4.5 exps per log: 96 us at 608
bits, 11x mpmath's pure-Python backend. It now reduces the way mpmath does
below 2500 bits: x = 2^e m with m in [sqrt(1/2), sqrt(2)), c = j / 512 the
nearest grid point, log x = e ln 2 + log c + 2 atanh(v) with
v = (m - c) / (m + c), |v| < 2^-10, so the series gains ~21 bits per term,
and a result near zero (x near 1) comes from the series alone. The table,
log(j / 512) = 2 atanh((j - 512) / (j + 512)) for j = 362 .. 724, uses
small-integer term updates only and is cached per thread at the widest
precision asked. Past 8192 working bits Newton stays.

Both log's atanh series and exp's Taylor series carry the running term --
and its multiplier and divisor, since a product or quotient takes the wider
precision -- at a precision that falls with the terms. The binary exponent
e is formed at the working precision: an efloat made from an integer has
the type's default precision, too narrow past 2^(nlimbs x 32).

Best of five, gcc -O2, against mpmath 1.3.0 (python backend):
  608 bits:  log 96 -> 12.3 us (1.41x), exp 20 -> 15.6 us (1.37x)
  2048 bits: log 1381 -> 116 us (1.07x), exp 222 -> 150 us (1.7x)
With a cold table over 2000 random arguments, log is 3.1x (608) and 6.2x
(2048 bits) faster than Newton; warm, 6.2x and 9.9x.

Ordinary results are bit-identical to main (12000 log/exp/near-1 samples at
320 and 608 bits); the #1712 accuracy suite passes at 320, 640 and 2048
bits. exp_log_precision.cpp now also checks log(2^e) for six exponents past
2^32, where main's Newton was off by 65 units or returned inf.

Resolves #1714

BEHAVIOUR-CHANGE: efloat log(x) for binary exponents beyond the type's default precision (e.g. |e| > 2^32 in efloat<1>) is now correctly rounded; it was off in the last places, or inf for tiny x

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

92 of 111 new or added lines in 2 files covered. (82.88%)

2 existing lines in 1 file now uncovered.

51422 of 58812 relevant lines covered (87.43%)

9338575.41 hits per line

Uncovered Changes

Lines Coverage ∆ File
19
76.44
0.39% include/sw/universal/number/efloat/math/logarithm.hpp

Coverage Regressions

Lines Coverage ∆ File
2
82.83
-0.3% include/sw/universal/number/qd/qd_impl.hpp
Jobs
ID Job ID Ran Files Coverage
1 39202799770.1 19 Oct 2026 05:25AM UTC 754
87.43
GitHub Action Run
Source Files on build 39202799770
  • Tree
  • List 754
  • Changed 4
  • Source Changed 2
  • Coverage Changed 4
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #39202799770
  • d852c650 on github
  • Prev Build on main (#39198997099)
  • Next Build on main (#39274580412)
STATUS · Troubleshooting · Open an Issue · Sales · Support · CAREERS · ENTERPRISE · START FREE TRIAL · SCHEDULE DEMO
ANNOUNCEMENTS · TWITTER · TOS & SLA · Supported CI Services · What's a CI service? · Automated Testing

© 2026 Coveralls, Inc