stillwater-sc / universal / 30683503007
85%
master: 84%

LAST BUILD BRANCH: main
DEFAULT BRANCH: master
Ran
Jobs 1
Files 600
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
coverage: 85.211% (-0.02%) from 85.234%
30683503007

push

github

web-flow 
fix(quire): size cfloat quire radix_point from smallest normal LSB (#1202) (#1288)

* fix(quire): size cfloat quire radix_point from smallest normal LSB (#1202)

quire_traits<cfloat>::radix_point was derived from |min value scale| in the
no-subnormals branch (bias-1), which is the scale of the smallest normal's
MSB (hidden bit). But that value carries fbits fraction bits below it, down
to 2^(1-bias-fbits), so its square needs 2*(bias+fbits-1) fraction bits.
The old no-subnormals radix_point (2*(bias-1)) was 2*fbits short, so the
quire truncated products of small-but-normal operands: an "exact" cfloat<16,5>
dot floored at ~2^-28 (rel err ~1e-9, worse than accumulating in double), and
cfloat<8,2> had radix_point 0 (no fraction bits at all).

The smallest representable magnitude's LSB is at scale (1-bias-fbits) in BOTH
subnormal modes -- without subnormals the smallest normal's fraction reaches
exactly as low as the smallest subnormal -- so radix_point does not depend on
hasSubnormals. Unify abs_min_scale = bias + fbits - 1 (guarded for the
degenerate bias+fbits==0 case), matching the native float/double
specializations and the file's own documented examples (cfloat<8,3>=12,
cfloat<32,8>=298, which the buggy code contradicted at 4 and 252).

Regression (static/quire/api/cfloat.cpp TestExactDotSmallNormals): an exact
quire dot over near-minpos NORMAL operands is bit-exact vs an independent
double oracle for cfloat<8,2> and cfloat<16,5> in BOTH subnormal modes, plus
static_asserts on the documented radix_point values. Fails on pre-fix headers
(cfloat<8,2> delivered 8 vs exact 10.37; cfloat<16,5> ~1e-9 off), passes with
the fix.

Verified gcc AND clang: full static/quire/api suite (quire_traits, cfloat,
cfloat_fdp, aggregator_fdp, api, fixpnt_fdp, lns_fdp) PASS; the larger
no-subnormals cfloat quire introduces no regressions.

Resolves #1202

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs(quire): refresh stale cfloat quire geom... (continued)

41699 of 48936 relevant lines covered (85.21%)

7271654.87 hits per line

Coverage Regressions

Lines Coverage ∆ File
11
31.07
 -3.93% include/sw/universal/verification/test_suite_randoms.hpp
Jobs
ID Job ID Ran Files Coverage
1 30683503007.1 600
85.21
 GitHub Action Run
Source Files on build 30683503007