hist VS awkward

Both projects are part of Scikit-HEP, but like many Scikit-HEP projects (Awkward, iminuit, cookie, etc.), are more generally applicable than just High Energy Physics (HEP). See https://github.com/scikit-hep/hist and https://github.com/scikit-hep/boost-histogram , drop us a star if this is useful for you!

there's a whole ecosystem in Python originally developed for high energy physics data processing: https://github.com/scikit-hep/awkward all because Numpy demands square N-dimensional array

Same technique used everywhere, here's a simple Julia pkg for the same thing: https://github.com/JuliaArrays/ArraysOfArrays.jl/blob/3a6f5b...

But Julia at least has the decency to just support ragged Vector{Vector} out of the box, and it's not that slow

Awkward Array.

Numba's @vectorize decorator (https://numba.pydata.org/numba-doc/latest/user/vectorize.htm...) makes a ufunc, and Awkward Array knows how to implicitly map ufuncs. (It is necessary to specify the signature in the @vectorize argument; otherwise, it won't be a true ufunc and Awkward won't recognize it.)

When Numba's JIT encounters a ctypes function, it goes to the ABI source and inserts a function pointer in the LLVM IR that it's generating. Unfortunately, that means that there is function-pointer indirection on each call, and whether that matters depends on how long-running the function is. If you mean that your assembly function is 0.1 ns per call or something, then yes, that function-pointer indirection is going to be the bottleneck. If you mean that your assembly function is 1 μs per call and that's fast, given what it does, then I think it would be alright.

If you need to remove the function-pointer indirection and still run on Awkward Arrays, there are other things we can do, but they're more involved. Ping me in a GitHub Issue or Discussion on https://github.com/scikit-hep/awkward-1.0