Accelerate xclim import

[aulemahal]

Generic Issue

• xclim version: all
• Python version: any
• Operating System: any

Description

Importing xclim.sdba is currently quite slow. I think this is due to the numba utilities being pre-compiled because we give explicit input types.

What I Did

In perfimport.py:

import xclim as xc
from xclim import sdba


if __name__ == '__main__':
    print('yeah')

and then:

python -X importtime perfimport.py 2> import.log
tuna import.log

Using tuna.

What I Received

I believe the other long import, "xclim/indices/_conversion.py", may be due to the UTCI function, which also uses numba.

Code of Conduct

• I agree to follow this project's Code of Conduct

[Zeitsperre]

Do you mean to say that this can be solved by relaxing static typing in SDBA? Is this a "feature" of numba?

[aulemahal]

Yep, this is what I read from the doc! https://numba.readthedocs.io/en/stable/user/jit.html#eager-compilation

IIRC, my goal with the static typing was avoid dtype promotions (float32 to 64). With this "eager" compilation, the tests would fail if the functions received mixed inputs, which I would then try to explicitly avoid. It may be thus be unnecessary to have them "in production". Also, a dtype promotion is not a tragic bug...