Evgeni Burovski

No activity since 2021, observed behavior is explained, recommended alternatives should indeed be more appropriate. Closing, thanks all.

> ... the scipy version was considerably slower than the numpy equivalent, ... The most notable one being the solve function where scipy is approximately 4x slower than numpy. Did...

Also, if you're benchmarking things, it would be interesting to see how scipy's structure detection fares. E.g. `solve` uses a Cholesky solver if it detects that the l.h.s. is PSD.

Curiously, a quick-and-dirty benchmark (under the fold) ``` import numpy as np from scipy.linalg import solve m, n = 500, 100 rng = np.random.default_rng(1234) a0 = rng.uniform(size=(m, n, n)) a...

> non-trivial time is spent in computing the condition number, at least with OpenBLAS Same with MKL: stubbing out the call to `pocon` saves about 50% of time on a...

> For example, one idea would be to re-write the Python wrapper for solve() in Cython, and do the argument validation / dispatching in Cython, and see if you can...

> Just pointing out that maybe we should expand on https://numpy.org/devdocs/reference/routines.linalg.html#linear-algebra and have some similar note in the SciPy docs? Absolutely yes please! On both counts, updating the numpy note,...

One example is https://github.com/scipy/scipy/pull/23919, which converts batched SVD from a python loop to C. For functions with multiple options (e.g. `eig`, qr etc), you don't need to convert them all...

> An example of a matrix that is not classified as being "banded" by this heuristic would be a Hessenberg matrix, You probably do not want to use a banded...

A partial and opinionated wishlist: > Do you allow C++ code, and if yes, in what standard? SciPy allows C++17, cf https://scipy.github.io/devdocs/dev/toolchain.html > Do you need a completely API-compatible alternative...