Not possible to use `eval_linear` with interpolation option ("xto.*") within jitted function about interpolation.py

[albop]

@gboehl Thanks for your interest.
The xto.LINEAR, xto.CONSTANT and other options are objects which are just designed so as to hold different types so multiple dispatch can occur. Initial idea was to define them as different jitted classes which are then automatically of a different type (see https://github.com/EconForge/interpolation.py/blob/master/interpolation/splines/option_types.py).
But, as I just learnt, jitclasses don't seem to be unboxed correctly in njitted context, numba/numba#2501 . I have pushed a temporary workaround where xto.LINEAR is just a random tuple value. The implementation is ugly but it seems to work. Could you try master a few minutes from now ?

[gboehl]

Thanks for the quick fix. It seems that you closed your PR instead of merging it. Anyways, I checked out albob/fix_xto and now using the xto.LINEAR option does not throw errors. Results seem reasonable.

Many thanks!

[albop]

Merged. Le dim. 11 août 2019 à 02:11, Gregor Boehl <[email protected]> a écrit :

…

Thanks for the quick fix. It seems that you closed your PR instead of merging it. Anyways, I checked out albob/fix_xto and now using the xto.LINEAR option does not throw errors. Results *seem* reasonable. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#52?email_source=notifications&email_token=AACDSKOJAL6UX6G52S4EM4DQD5KKPA5CNFSM4IKXPVCKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD4AXKTY#issuecomment-520189263>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AACDSKODMJLGOSKXU5G3VQ3QD5KKPANCNFSM4IKXPVCA> .

[gboehl]

Sorry, this is not a fix as it seems to be messing with the shape of the output of the interpolation. I'll have closer look next week.

[albop]

Strange, I would have assume output shape would be independent from the interpolation method. I don't have any obvious anomaly on my laptop. If you want to look I would suggest to start with splines/eval_splines.py . This is were the time dispatch is implemented. It is a bit heavy because of the need to overload function. Dispatch rules determine the source code of the actual spline interpolation logic. You can just look at the output of lines 102 and 127 to see exactly what is being generated (maybe we should have a debug option somewhere instead of commented print statements...)

[gboehl]

I can't reproduce this on my workstation with the latest versions of the related packages, so I assume that the error was somehow related to my python environment. Thanks a lot for the quick reaction!

(closing)