Module stretch 

Standard-deviation contrast stretch, as a fused single-pass kernel.

Mirrors TerraTexture.stretch’s stretch_std(): values are mapped linearly so that mean - n_std·std becomes 0 and mean + n_std·std becomes 1, then clipped to [0, 1]. NaN in means NaN out, and NaNs are excluded from the mean and standard deviation.

Why this is faster than numpy

The numpy version calls np.nanmean() and then np.nanstd(), which recomputes the mean internally, so the array is reduced several times before the elementwise subtract/divide/clip. Here one reduction pass collects the sum, sum of squares and count together. (Deduplicating only the mean call in numpy gives about 1.1×; numpy’s C reductions are already efficient, so the win is fewer passes, not language speed.)

Precision

Statistics are accumulated in f64 for robustness on large arrays. This deliberately does not bit-match numpy’s float32 pairwise summation; it is verified against it within a statistical tolerance in tests/test_stretch_rust.py, which is the right bar for a mean/std computation.

Functions

stretch_std_core

Contrast-stretch an array to [0, 1] using mean ± n_std standard deviations, ignoring NaNs.