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

stillwater-sc / universal / 43841038115
89%
master: 84%

Build:
Build:
LAST BUILD BRANCH: fix/fixpnt-quire-resolve-saturates
DEFAULT BRANCH: master
Ran 16 Nov 2026 09:52AM UTC
Jobs 1
Files 897
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

16 Nov 2026 09:34AM UTC coverage: 89.214% (+0.02%) from 89.192%
43841038115

push

github

web-flow
fix(microfloat): mxfp8e5m2 clips, max_exponent, one rounding from double and integers, ++/-- at -0 and maxpos; nvblock NaN input (#2102)

- mxfloat: mxfp8e5m2 is mxblock<e5m2_saturating>, a new element type with
  e5m2's encoding that clamps on overflow, as #1302 did for mxfp8. The scaled
  block maximum lies in [2^15, 2^16) against a maxpos of 57344, and every value
  from 61440 up was inf. An infinite input still gives inf.
- microfloat: numeric_limits::max_exponent is maxpos's scale + 1: e4m3 reported
  8 for 9, e2m1 2 for 3, e2m3 2 for 3, e3m2 4 for 5.
- microfloat: double, long double and integers are rounded once. They were cast
  to float first, which rounds to nearest before the microfloat rounds again:
  e4m3(1.0625 + 1e-12) was 1.0, not 1.125. The source is narrowed to float with
  round-to-odd, which keeps two more bits than any microfloat significand.
- microfloat: ++(-0) is +minpos, and ++ / -- stop at maxpos / maxneg and leave
  inf and NaN where they are. ++ from e5m2's maxpos 0x7B gave inf.
- nvblock: a NaN input sets the block scale to NaN, as #2086 did for MX blocks,
  so the whole block dequantizes to NaN. It was dropped: the scale came from the
  other elements and e2m1 turned the NaN into 0. The division by the tensor scale
  is done once per element, not twice.
- microfloat: narrowing a double or long double keeps the sign of -0.
- mxblock: byte_size() documents that it is this implementation's storage, one
  byte per element, not OCP's packed size.

Per #2096: an inf input to an MX block keeps today's behaviour (now pinned),
and microfloat's integer casts are unchanged.

Tests: microfloat/conversion/rounding_and_stepping.cpp (max_exponent against
maxpos, near ties from double and long double, integers past 2^24, ++ / -- over
every encoding); nvblock/conversion/special_inputs.cpp (NaN, inf, tensor scale
0, tiny amax); mxfloat/conversion/ocp_scaling.cpp (the block maximum clips,
inf input, byte_size).

Resolves #2096

BEHAVIOUR-CH... (continued)

56 of 56 new or added lines in 3 files covered. (100.0%)

3 existing lines in 1 file now uncovered.

64046 of 71789 relevant lines covered (89.21%)

12414162.58 hits per line

Coverage Regressions

Lines Coverage ∆ File
3
34.38
-0.95% include/sw/universal/verification/test_suite_randoms.hpp
Jobs
ID Job ID Ran Files Coverage
1 43841038115.1 16 Nov 2026 09:52AM UTC 897
89.21
GitHub Action Run
Source Files on build 43841038115
  • Tree
  • List 897
  • Changed 8
  • Source Changed 6
  • Coverage Changed 7
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #43841038115
  • 3d37d894 on github
  • Prev Build on main (#43834366987)
  • Next Build on main (#43915301650)
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