Help needed! How to specify numba.jit arguments for read-only numpy arrays?

[Jason Sachs]

I am in the process of upgrading some old code from numba 0.16.0 (!!!) to numba 0.33.0 and it looks like one aspect of my old code is broken; I get a TypeError because one of the arguments to a function is a read-only numpy array, and it's incompatible with the numba.jit type signature I gave it.

How do you specify a type is read-only?

Here's a short example that illustrates the problem; this worked fine on numba 0.16 but gives a TypeError on numba 0.33:

-----

import numba

import numpy as np

@numba.jit('f8(u4[:])')

def jabberwocky(data):

    x = 0.0

    w = 1

    for item in data:

        x += 1.0*(item*w)

        w = -w

    return x

    

data = np.array([1,2,3,4,5],dtype=np.uint32)

data.flags.writeable = False

print data

print jabberwocky(data)

-----

> python numba_test1.py

[1 2 3 4 5]

Traceback (most recent call last):

  File "numba_test1.py", line 16, in <module>

    print jabberwocky(data)

  File "\path\to\anaconda\envs\myenv\lib\site-packages\numba\dispatcher.py", line 399, in _explain_matching_error

    raise TypeError(msg)

TypeError: No matching definition for argument type(s) readonly array(uint32, 1d, C)

-----

Is there a way to fix?

[Jason Sachs]

I figured out an option after studying http://numba.pydata.org/numba-doc/dev/reference/types.html and http://numba.pydata.org/numba-doc/dev/reference/jit-compilation.html

This works in both 0.16.0 and 0.33.0, using numba type objects rather than a string signature:

-----

import numba

import numpy as np

x0 = np.empty((0,0), dtype=np.uint32)

x0.flags.writeable = False

@numba.jit(numba.double(numba.typeof(x0)))

def jabberwocky(data):

    x = 0.0

    w = 1

    for item in data:

        x += 1.0*(item*w)

        w = -w

    return x

    

data = np.array([1,2,3,4,5],dtype=np.uint32)

data.flags.writeable = False

print data

print jabberwocky(data)

----

Is there any way to avoid making this dummy object x0, and somehow create a type object directly that specifies read-only?

[Jason Sachs]

oops that should have been:

x0 = np.empty((0), dtype=np.uint32)

odd how that still worked... :/