DimShuffle's C code does not correctly handle ndarray views

[brandonwillard]

Here's a simple example:

import aesara.tensor as at


at.broadcast_to(0, (2, 3))[None].eval()
# array([[[  0, 122, -76],
#         [-88,  49,  86]]], dtype=int8)

Here's the same result, just computed more directly via the generated C thunk:

import numpy as np
import aesara


x_at = at.vector()
x_fn = aesara.function([x_at], x_at[None])

x_fn.fn.fgraph.toposort()
# [InplaceDimShuffle{x,0}(<TensorType(float64, vector)>),
#  DeepCopyOp(InplaceDimShuffle{x,0}.0)]

dimshuffle_thunk = x_fn.fn.thunks[0]

dimshuffle_thunk
# <function aesara.graph.op.COp.make_c_thunk.<locals>.rval()>

dimshuffle_thunk.inputs[0][0] = np.broadcast_to(0, (6,))
dimshuffle_thunk()
dimshuffle_thunk.outputs[0][0]
# array([[              0,               0, 140009315547216,
#                      33,               0,               0]])

np.shares_memory(dimshuffle_thunk.inputs[0][0], dimshuffle_thunk.outputs[0][0])
# True

[brandonwillard]

It looks like the C implementation attempts to change the dimensions and strides itself, and it probably makes some strong assumptions about them in the process. Since broadcasting changes that information, those assumptions are probably bad.

Instead, we should be able to use PyArray_Transpose and PyArray_Newshape in the same way the Python implementation does.