This is a Rust version of nnls (Non-Negative Least Squares). It's a port from a Fortran90 script used by scipy.optimize.nnls and probably several others.
I decided to port nnls to Rust because:
- The C version of nnls is not thread-safe! There are several
staticvariables throughout the code which make it unusable in parallel. If you plan to use nnls only in single-thread, you might prefer linking to that version: it has been heavily tested and has no known problem. - Both Fortran versions (77 and 90) require some knowledge of Fortran memory model and some compiler tricks to successfuly link against. I wasn't able to do it and I didn't want to waste more hours into this adventure, so I simply ported the script to Rust.
- It has been used several millions of times and it will be used billions of times soon. I work with 3D images and each image I see contains around 1 million voxels. For a specific algorithm, I need to call
nnls2 times for each voxel. Several of those images has been tested and compared with the original Fortran version. - There's no
unsafeand it is forbidden to use any in this crate.
- This is not idiomatic Rust. I tried cleaning and enhancing the code but it's not always possible.
- I use ndarray. You might prefer something else. I do not plan to change this but you're welcome to discuss it.
- There are currently 1
panic!in the code because there are 1 code paths in the original code which usegotothat I couldn't translate properly to Rust. I planned to handle that case but it turns out that it's never called. As I wrote,nnlshas been extensively used and I couldn't find any dataset that triggers that condition. If my programs ever panic because of this, it will be repaired quickly.