Set up interface between Uproot and Awkward so that Awkward can be used to optimize object-reading.

[jpivarski]

As discussed in issue #90.

Corresponds to scikit-hep/awkward#448.

[jpivarski]

@tamasgal The good news is that, with this PR and scikit-hep/awkward#448, the following code takes 3.6 seconds instead of 194 seconds (53× faster):

import uproot4
branch = uproot4.open("issue-90.root:E/Evt/trks/trks.fitinf")
for i in range(branch.num_baskets):
    print(repr(branch.basket(i).array()))

In the above, we're reading each TBasket individually. The bad news is that it still takes a long time to concatenate TBaskets because ak.concatenate has only been (internally) implemented for pairs: concatenating n arrays means creating and throwing away n - 2 temporary arrays of quadratically increasing size. That was fine for examples of concatenating 2 or 3 arrays at a time, but when it comes to concatenating data from this file's 461 TBaskets, it's a problem.

Clearly, ak.concatenate needs to be fixed anyway. There might already be an issue open about it. (It's been on my mind for a while...) Anyway, I'll tackle that next. We want the conclusion of the above story to be that the whole array is produced in 3.7 seconds!

[tamasgal]

Awesome! Many thanks, Jim, that's really a huge leap.

[jpivarski]

For cases that aren't covered by the new interpret-by-Awkward mechanism, how about parallelizing the pure Python interpretation?

• parallel interpretation in 'basket_array': 259 seconds. 12 cores utilized at about 5-10% each. I'll blame the GIL. Reversing the order (putting the big TBaskets first) doesn't help—it's not about stragglers.
• sequential interpretation in 'basket_array', concatenate NumPy dtype=O arrays: 190 seconds.
• sequential interpretation in 'final_array' (old way): 197 seconds. Moving the interpretation from final_array into basket_array doesn't hurt. Entries that would be trimmed (because they're at the ends of the first and last basket) are now unnecessarily interpreted, but that's probably better than not having the possibility to parallelize.
• interpret-by-Awkward in 'basket_array': 2.5 seconds.

So the baseline of 197 seconds wasn't unnecessarily harsh. It really is an 80× speedup.

[jpivarski]

Making use of this will require Awkward 0.2.37 (but it won't break with earlier versions), which is being deployed now. I'll do one last test after that deployment so that GHA pulls the new Awkward from PyPI.

[tamasgal]

Thanks, that's really nice and helps a lot!