scikit-hep / iminuit / 34977343608

100%

push

github

perf: store LeastSquares data in Fortran order for contiguous columns (#1176)

:robot: _AI text below_ :robot:

`LeastSquares` passed strided column views of the `(N, ndim + 2)` data
array to the model and the cost on every evaluation. The data array is
now built with `np.stack(...).T`, which makes it Fortran-ordered, so the
columns are contiguous views. Masked data is made Fortran-ordered again
in `_update_cache`, because fancy indexing returns a C-ordered copy.

At N = 200k, the 1D `__call__` is 17% faster and the 2D `__call__` is
26% faster. Construction is about 3x faster as well, because `np.stack`
writes one contiguous array instead of the strided copy that
`np.column_stack` made.

Two new tests cover the `data` setter and 2D data with a mask. Both fail
if the cache is stale.

11 of 11 new or added lines in 1 file covered. (100.0%)

3991 of 4002 relevant lines covered (99.73%)

1.0 hits per line