nsajko

First point: Parameter range is defined in the Sollya script that is supposed to be saved to /tmp/polynomials.sollya. This is the one generated by print_poly_from_coefficients.cc, so that's where the definition is (lines 115 to 123): https://gitlab.com/nsajko/geodesic_experiments_for_geographiclib/-/blob/main/src_print_poly/print_poly_from_coefficients.cc#L115 So there is dom for A1 and A2, and dom2 for C1, C1p, C2. The range dom is defined again for the random generator functions, as discussed in...

One thing that I should have noted somewhere in the README on the Gitlab repo is that the C++ code in that repo is C++20, tested with GCC 11. And I know for a fact that it uses multiple features which are not yet supported by Clang.

First point: Parameter range is defined in the Sollya script that is supposed to be saved to /tmp/polynomials.sollya. This is the one generated by print_poly_from_coefficients.cc, so that's where the definition is (lines 115 to 123): https://gitlab.com/nsajko/geodesic_experiments_for_geographiclib/-/blob/main/src_print_poly/print_poly_from_coefficients.cc#L115 So there is dom for A1 and A2, and dom2 for C1, C1p, C2. The range dom is defined again for the random generator functions, as discussed in...

As noted in the repo linked from the commit message, there is a sort of issue with the rand0 relative error measurements: I don't understand how such results are possible (for rand0 only). Another issue that requires more thinking about is that this change set optimizes for WGS84, so it's no good for ellipsoids very different from it. So maybe you'd want to guard this changeset with a new macro, perhaps named GEOGRAPHICLIB_OPTIMIZE_FOR_WGS84.

Max relative errot plots (see the URL in the commit message):

Geodesic: improve A1m1f, A2m1f, C1f, C1pf, C2f

Clang, on the other hand, also has that option, but with a different default.