Build: #156 was successful Changes by awskii <awskii@users.noreply.github.com>
Code commits
GitHub Casacore
-
awskii <awskii@users.noreply.github.com> 3c16927b905dc3e690ad8dfe944afd58042bc447
Fix LSQFit chi2 cancellation that fails tLSQ tests under FMA (#1548)
* Compute LSQFit chi2 as yy - 2x.b + x.N.x so FMA cannot inflate exact-fit SDs (#1367, #1352)
* Halve the LSQFit diagonal product, not the diagonal, and add SUMLL after the quadratic sum; test both edge cases
* Mask tLSQ exact-fit prints at 1e-5: chi2 floor is one ulp of yy, SD 7.67e-6 on all five builds
* Compute LSQFit chi2 as SUMLL - x.b + x.(N.x - b) without halving: overflow, cancelling and subnormal cases match master
* Fall back to master's LSQFit chi2 sum when the second-order term overflows, so solutions near DBL_MAX keep master's SD
* Sum LSQFit constraint terms only inside the chi2 overflow fallback, so normal fits run 993dacaf7's path unchanged
* Update scimath/Fitting/LSQFit.cc
---------
Co-authored-by: André Offringa <offringa@gmail.com>- .github/workflows/osx.yml (version 3c16927b905dc3e690ad8dfe944afd58042bc447)
- scimath/Fitting/LSQFit.cc (version 3c16927b905dc3e690ad8dfe944afd58042bc447)
- scimath/Fitting/test/tLSQFit.cc (version 3c16927b905dc3e690ad8dfe944afd58042bc447)
- scimath/Fitting/test/tLSQFit.out (version 3c16927b905dc3e690ad8dfe944afd58042bc447)
- scimath/Fitting/test/tLSQaips.cc (version 3c16927b905dc3e690ad8dfe944afd58042bc447)