Numba gufuncs slowing down import (due to eager mode compilation)

[frazane]

Importing the library seems to be significantly slowed down by numba, see output of

PYTHONVERBOSE=1 python -c "from scoringrules.core import energy"

This is likely related to how guvectorize works. It has two modes and we're using "eager, decoration-time compilation". In other words, we're compiling functions during import.

Eager, or decoration-time, compilation: If you pass one or more type signatures to the decorator, you will be building a NumPy universal function (ufunc). The rest of this subsection describes building ufuncs using decoration-time compilation.

Lazy, or call-time, compilation: When not given any signatures, the decorator will give you a Numba dynamic universal function (DUFunc) that dynamically compiles a new kernel when called with a previously unsupported input type. A later subsection, “Dynamic universal functions”, describes this mode in more depth.

Potential solutions?

• Use lazy, call-time compilation
• Use lazy imports of numba modules
• Use caching

[frazane]

After some investigation there also seem to be an issue when using np.linalg.norm inside of numba gufuncs. Substituting it with a custom jitted function improves import time considerably (not sure about performance impact).

[frazane]

Could be related to #68 as we could have fewer numba gufuncs

[frazane]

Note to self: caching could help.