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

stillwater-sc / universal / 43721160556
89%
master: 84%

Build:
Build:
LAST BUILD BRANCH: fix/fixpnt-quire-resolve-saturates
DEFAULT BRANCH: master
Ran 15 Nov 2026 02:08PM UTC
Jobs 1
Files 894
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

15 Nov 2026 01:51PM UTC coverage: 89.14% (+0.01%) from 89.13%
43721160556

push

github

web-flow
fix(mxblock): mxint8 emax 6 and ties to even; exact shared exponent, clamped scale; inline max_elem_exponent (#2085)

mxblock::quantize did not follow OCP MX v1.0 section 6.3 in four ways (#2083),
and in one more found while fixing them:
- max_elem_exponent<int8_t> was 7. An int8 code's largest power of two is
  2^6 (OCP's INT8 has an implicit 2^-6, emax 0), so every block's largest
  code fell in [128, 256) and was clamped to 127: {100, 1.5, 2.5, -3.25}
  came back {63.5, 1.5, 2.5, -3.5}. It is 6, and the block round-trips to
  {100, 2, 2, -3}.
- int8 codes rounded with std::round, ties away from zero (2.5 -> 3). They
  round to nearest, ties to even, by one constexpr routine for both paths.
- The max_elem_exponent specializations were not inline, so two translation
  units including mxfloat.hpp failed to link.
- floor(log2(amax)) was clamped, not the scale exponent, and the elements
  were scaled by a reciprocal that is inf below 2^-127: mxfp8 {2^-126, 0, 0}
  became {2.6e-36, NaN, NaN}. The scale exponent is clamped to e8m0's range
  and the elements are scaled with ldexp.
- floor(log2(amax)) came from a float std::log2, which rounds up just below
  a power of two (16 - 2^-20 gave shared exponent 4, not 3) and casts inf to
  int for an infinite amax. It comes from std::ilogb, exact. An infinite amax
  takes shared exponent 127 at runtime and at constant evaluation alike.
- An all-zero block stored -0 as +0; its elements are converted like any
  other block's now.

Tests:
- mxfloat/conversion/ocp_scaling: every scale and element against a double
  reference with its own emax table and std::frexp, for all six aliases:
  blocks just below and at every power of two, the bottom and top of the
  scale range, and random blocks. Plus the issue's blocks, int8 ties over
  [-64, 64), and 192 int8 block maxima. On main 1221 to 5195 cases fail per
  alias, and 189 of 192 maxima are clipped.
- mxfloat/api/constexpr: the same at constant evaluation (emax, ties, t... (continued)

27 of 29 new or added lines in 3 files covered. (93.1%)

63435 of 71163 relevant lines covered (89.14%)

12007889.91 hits per line

Uncovered Changes

Lines Coverage ∆ File
2
65.08
0.79% integration/multifile/main.cpp
Jobs
ID Job ID Ran Files Coverage
1 43721160556.1 15 Nov 2026 02:08PM UTC 894
89.14
GitHub Action Run
Source Files on build 43721160556
  • Tree
  • List 894
  • Changed 5
  • Source Changed 2
  • Coverage Changed 5
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #43721160556
  • 6e05d9f9 on github
  • Prev Build on main (#43684754438)
  • Next Build on main (#43749674577)
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