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

stillwater-sc / universal / 39047295777
89%
master: 84%

Build:
Build:
LAST BUILD BRANCH: fix/issue-2132-rational-exact-saturating
DEFAULT BRANCH: master
Ran 17 Oct 2026 04:15PM UTC
Jobs 1
Files 725
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

17 Oct 2026 03:57PM UTC coverage: 86.692% (+0.01%) from 86.68%
39047295777

push

github

web-flow
feat(lindex): shared phi/psi kernel and mpmath oracle (#1349 part 2) (#1707)

* feat(lindex): shared phi/psi kernel and mpmath oracle (#1349 part 2)

number/shared/level_index_kernel.hpp is the level-index kernel that lindex
(#1694) and slindex (#1350) build on. It provides:
- phi and psi;
- the exposed Clenshaw-Olver a and b sequences, for SLI to reuse;
- psi-domain add and sub.
It never forms a value. A psi is a point { level, index }, so the index
stays exact in any working type W wide enough for the format's index bits.
Subtraction carries d_j = 1 - b_j through expm1 when the operands share a
level or sit in adjacent levels; its start, -expm1(g - f) or
-expm1((g - 1) - expm1(f)), is exact. x - x is decided before the
sequences. The kernel includes no I/O.

tools/generators/level_index_gen.py is the oracle. mpmath evaluates the
plain sequences at 600 bits, which is a different formulation from the
kernel's expm1 form. Wherever psi <= 4.6 it cross-checks them against
direct evaluation in the value domain, to 2^-200 (316 of the 592 pairs at
fbits 28). It emits encodings and exact psi as hex double limbs into
verification/level_index_vectors.hpp, and the output reproduces byte for
byte. An a_j below 2^-1500 is taken as 0, the limit the kernel also takes;
the generator bounds the error that introduces.

static/logarithmic/lindex/kernel/kernel.cpp measures error in psi-ulps over
592 pairs per format: every level, near cancellation, level boundaries,
y = 0 and x - x.
- double at fbits 28: worst 3.0e-7 ulp. dd_cascade at fbits 60: 3e-14.
- Every result is faithful and correctly rounded, and x - x is 0.
- double at fbits 60 is reported only: 132 results are not faithful, which
  confirms the design doc's choice of dd_cascade above 32 bits.
The bounds sit about 50x above the measurements. All six kernel mutants
tried are caught. The measured table goes into the design doc.

CMake: a new UNIVERSAL_BUILD_NUMBER_LINDEX option builds
static/logarithmic/lindex, ... (continued)

50 of 50 new or added lines in 1 file covered. (100.0%)

49482 of 57078 relevant lines covered (86.69%)

10424427.06 hits per line

Jobs
ID Job ID Ran Files Coverage
1 39047295777.1 17 Oct 2026 04:15PM UTC 725
86.69
GitHub Action Run
Source Files on build 39047295777
  • Tree
  • List 725
  • Changed 0
  • Source Changed 0
  • Coverage Changed 0
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #39047295777
  • fced16af on github
  • Prev Build on main (#38995555605)
  • Next Build on main (#39053590561)
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