Awkward Array features that are supported in Numba-compiled functions#

See the Numba documentation, which maintains lists of

in JIT-compiled functions. This page describes the supported Awkward Array library features.

import awkward as ak
import numpy as np
import numba as nb

Passing Awkward Arrays as arguments to a function#

The main use is to pass an Awkward Array into a function that has been JIT-compiled by Numba. As many arguments as you want can be Awkward Arrays, and they don’t have to have the same length or shape.

array1 = ak.Array([[0, 1.1, 2.2], [], [3.3, 4.4], [5.5], [6.6, 7.7, 8.8, 9.9]])
array2 = ak.Array([
    [{"x": 1.1, "y": [1]}, {"x": 2.2, "y": [1, 2]}, {"x": 3.3, "y": [1, 2, 3]}],
    [],
    [{"x": 4.4, "y": [1, 2, 3, 4]}, {"x": 5.5, "y": [1, 2, 3, 4, 5]}]
])
@nb.jit
def first_array(array):
    for i, list_of_numbers in enumerate(array):
        for x in list_of_numbers:
            if x == 3.3:
                return i

@nb.jit
def second_array(array):
    for i, list_of_records in enumerate(array):
        for record in list_of_records:
            if record.x == 3.3:
                return i

@nb.jit
def where_is_3_point_3(a, b):
    return first_array(a), second_array(b)
where_is_3_point_3(array1, array2)
(2, 0)

The only constraint is that union types can’t be accessed within the compiled function. (Heterogeneous parts of an array can be ignored and passed through a compiled function.)

Returning Awkward Arrays from a function#

Parts of the input array can be returned from a compiled function.

@nb.jit
def first_array(array):
    for list_of_numbers in array:
        for x in list_of_numbers:
            if x == 3.3:
                return list_of_numbers

@nb.jit
def second_array(array):
    for list_of_records in array:
        for record in list_of_records:
            if record.x == 3.3:
                return record

@nb.jit
def find_3_point_3(a, b):
    return first_array(a), second_array(b)
found_a, found_b = find_3_point_3(array1, array2)
found_a
[3.3,
 4.4]
-----------------
backend: cpu
nbytes: 16 B
type: 2 * float64
found_b
{x: 3.3,
 y: [1, 2, 3]}
--------------------------------------------
backend: cpu
nbytes: 192 B
type: {
    x: float64,
    y: var * int64
}

Cannot use ak.* functions or ufuncs#

Outside of a compiled function, Awkward’s vectorized ak.* functions and NumPy’s universal functions (ufuncs) should be highly preferred over for-loop iteration because they are much faster.

Inside of a compiled function, however, they can’t be used at all. Use for-loops and if-statements instead.

This is an either-or choice at the boundary of a @nb.jit-compiled function. (Even if ak.* had been implemented in Numba’s compiled context, it would be slower than compiled for-loops and if-statements because of the intermediate arrays they would necessarily create.)

Cannot use fancy slicing#

Similarly, any slicing other than

  • a single integer, like array[i] where i is an integer, or

  • a single record field as a constant, literal string, like array["x"] or array.x,

is not allowed. Unpack the data structures one level at a time.

Casting one-dimensional arrays as NumPy#

One-dimensional Awkward Arrays of numbers, which are completely equivalent to NumPy arrays, can be cast as NumPy arrays within the compiled function.

@nb.jit
def return_last_y_list_squared(array):
    y_list_squared = None
    for list_of_records in array:
        for record in list_of_records:
            y_list_squared = np.asarray(record.y)**2
    return y_list_squared
return_last_y_list_squared(array2)
array([ 1,  4,  9, 16, 25])

This ability to cast Awkward Arrays as NumPy arrays, and then use NumPy’s ufuncs or fancy slicing, softens the law against vectorized functions in the compiled context. (However, making intermediate NumPy arrays is just as bad as making intermediate Awkward Arrays.

Creating new arrays with ak.ArrayBuilder#

Numba can create NumPy arrays inside a compiled function and return them as NumPy arrays in Python, but Awkward Arrays are more complex and this is not possible. (Aside from implementation, what would be the interface? Data in Numba’s compiled context must be fully typed, and Awkward Array types are complex.)

Instead, arrays can be built with ak.ArrayBuilder, which can be used in compiled contexts and discovers type dynamically. Each ak.ArrayBuilder must be instantiated outside of a compiled function and passed in, and then its ak.ArrayBuilder.snapshot() (which creates the ak.Array) must be called outside of the compiled function, like this:

@nb.jit
def create_ragged_array(builder, n):
    for i in range(n):
        builder.begin_list()
        for j in range(i):
            builder.integer(j)
        builder.end_list()
    return builder
builder = ak.ArrayBuilder()

create_ragged_array(builder, 10)

array = builder.snapshot()

array
[[],
 [0],
 [0, 1],
 [0, 1, 2],
 [0, 1, 2, 3],
 [0, 1, 2, 3, 4],
 [0, 1, 2, 3, 4, 5],
 [0, 1, 2, 3, 4, 5, 6],
 [0, 1, 2, 3, 4, 5, 6, 7],
 [0, 1, 2, 3, 4, 5, 6, 7, 8]]
-----------------------------
backend: cpu
nbytes: 448 B
type: 10 * var * int64

or, more succintly,

create_ragged_array(ak.ArrayBuilder(), 10).snapshot()
[[],
 [0],
 [0, 1],
 [0, 1, 2],
 [0, 1, 2, 3],
 [0, 1, 2, 3, 4],
 [0, 1, 2, 3, 4, 5],
 [0, 1, 2, 3, 4, 5, 6],
 [0, 1, 2, 3, 4, 5, 6, 7],
 [0, 1, 2, 3, 4, 5, 6, 7, 8]]
-----------------------------
backend: cpu
nbytes: 448 B
type: 10 * var * int64

Note that we didn’t need to specify that the type of the data would be var * int64; this was determined by the way that ak.ArrayBuilder was called: ak.ArrayBuilder.integer() was only ever called between ak.ArrayBuilder.begin_list() and ak.ArrayBuilder.end_list(), and hence the type is var * int64.

Note that ak.ArrayBuilder can be used outside of compiled functions, too, so it can be tested interactively:

with builder.record():
    builder.field("x").real(3.14)
    with builder.field("y").list():
        builder.string("one")
        builder.string("two")
        builder.string("three")
builder.snapshot()
[[],
 [0],
 [0, 1],
 [0, 1, 2],
 [0, 1, 2, 3],
 [0, 1, 2, 3, 4],
 [0, 1, 2, 3, 4, 5],
 [0, 1, 2, 3, 4, 5, 6],
 [0, 1, 2, 3, 4, 5, 6, 7],
 [0, 1, 2, 3, 4, 5, 6, 7, 8],
 {x: 3.14, y: ['one', 'two', 'three']}]
--------------------------------------------------------------------------------------------
backend: cpu
nbytes: 614 B
type: 11 * union[
    var * int64,
    {
        x: float64,
        y: var * string
    }
]

But the context managers, with builder.record() and with builder.list(), don’t work in Numba-compiled functions because Numba does not yet support it as a language feature.

Overriding behavior with ak.behavior#

Just as behaviors can be customized for Awkward Arrays in general, they can be customized in the compiled context as well. See the last section of the ak.behavior reference for details.