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

stillwater-sc / universal / 37364685012
89%
master: 84%

Build:
Build:
LAST BUILD BRANCH: fix/issue-2131-integer-conversions-shift
DEFAULT BRANCH: master
Ran 05 Oct 2026 08:19PM UTC
Jobs 1
Files 688
Run time 2min
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

05 Oct 2026 07:38PM UTC coverage: 86.204% (+0.02%) from 86.183%
37364685012

push

github

web-flow
feat(bposit): exact fused dot products through the generalized quire (#1635)

quire_traits<bposit<nbits, rs, es, bt>> sizes the generalized quire so that
every product of two bposits lands in it exactly, and bposit/fdp.hpp adds
quire_mul, quire_resolve, fdp, fdp_stride and fdp_qc, mirroring posit's.

The layout follows cfloat's rule: the smallest product BIT at bit 0,
maxpos^2 below the top. A b-posit's smallest values still carry F_min
fraction bits, so the smallest product bit is 2^-(2 (rs 2^es + F_min)), not
minpos^2's scale:
  radix_point = 2 (rs 2^es + F_min), upper_range = 2 rs 2^es,
  range = 4 rs 2^es + 2 F_min
The exact quire therefore depends on n: 776 / 808 / 872 bits for
bposit16 / 32 / 64. The often-quoted "800 bits for any n > 12" matches only
n = 16; anchored at minpos^2 it drops the lowest 2 F_min bits of products
of tiny values (104 at n = 64) -- the trap #1202 fixed for cfloat. The
ranges were confirmed by brute force over all products of five small
configurations. half_range is the larger half, so quire::operator+= (which
skips out-of-window operands) accepts every product scale.

quire_resolve rounds once through bposit's own encoder, so a tiny or huge
sum saturates to +/-minpos or +/-maxpos, as every bposit result does,
rather than to zero.

Tests (static/tapered/bposit/arithmetic/fdp.cpp), against an exact
1024-bit integer reference -- every value and product is an integer at the
quire's scale: random dot products of length 1..48 for bposit<8,4,0>,
<12,4,2>, bposit16/32/64, checked for correct rounding, ties to even and
saturation; exact cancellation; NaR, zero and maxpos^2 cases; 1000
maxpos^2 terms added and subtracted; fdp_stride; and a*b - c*d near minpos,
nonzero only in the low bits, which must resolve to minpos at n = 16, 32
and 64. A mutation check confirms the suite catches a quire sized from
minpos^2 and a resolve without its sticky bit. Clean under ASan and UBSan;
the existing quire suites still pass.

Docs: a quire se... (continued)

36 of 36 new or added lines in 1 file covered. (100.0%)

2 existing lines in 1 file now uncovered.

46676 of 54146 relevant lines covered (86.2%)

9780317.41 hits per line

Coverage Regressions

Lines Coverage ∆ File
2
95.73
0.39% include/sw/universal/number/cfloat/cfloat_impl.hpp
Jobs
ID Job ID Ran Files Coverage
1 37364685012.1 05 Oct 2026 08:19PM UTC 688
86.2
GitHub Action Run
Source Files on build 37364685012
  • Tree
  • List 688
  • Changed 2
  • Source Changed 0
  • Coverage Changed 2
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #37364685012
  • 62850f20 on github
  • Prev Build on main (#37336179188)
  • Next Build on main (#37687696635)
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