You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fix spurious make check failures on the intensive driver tests (#78): the non-portable driver-vs-reference text comparison was dropped. Driver validation is now a regression gate against the original FORTRAN MINPACK — cminpack (pure C and f2c) must converge on every driver problem the committed reference (examples/ref/*.fortran.ref) converges on, so no Fortran compiler is needed. Iteration-count and last-digit differences (compiler FMA) never fail the build.
Add user-maintained exclusion lists for the few deliberately-extreme coin-flip problems (FORTRAN itself converges on them only at some optimization levels): examples/crosscheck_exclude.txt, plus examples/crosscheck_exclude_blas.txt for USE_BLAS/USE_LAPACK builds. Each gate failure prints the problem as nprob/dim, the token to add.
Replace crosscheck.sh with crosscheck.py; add driver_check.py (--reference-gate) and crosscheck_matrix.sh (an optimization-level diagnostic). Wire the cross-check into CMake/CTest (CMINPACK_CROSSCHECK); Python 3 is optional (only the cross-check needs it).
Fix a compile error in examples/tenorm_.c (pass dpmpar_ its index by pointer, the FORTRAN/f2c calling convention).
Document the numerical differences from FORTRAN MINPACK and correct the enorm.c comments (the divergence is FMA, not the rdwarf/rgiant constants).