cupy.lib.stride_tricks.as_strided(...) not consistent with Numpy?

[kkotyk]

Using strides causes memory allocations to balloon in size. It looks like as_strided() is not creating a view of the array the same way that Numpy does.

import cupy as cp
import numpy as np

def windowed_view(ndarray, window_len, step, xp):

    nrows = ((ndarray.shape[-1] - window_len) // step) + 1
    last_dim_stride = ndarray.strides[-1]
    new_shape = ndarray.shape[:-1] + (nrows, window_len)
    new_strides = list(ndarray.strides + (last_dim_stride,))
    new_strides[-2] *= step

    return xp.lib.stride_tricks.as_strided(ndarray, shape=new_shape, strides=new_strides)


arr = np.ones((1,1,1000))
v = windowed_view(arr, 100, 10, np)

print(v.shape)
print(arr.nbytes, v.nbytes)
print(v.base)
print(v.base.base.nbytes)

arr = cp.ones((1,1,1000))
v = windowed_view(arr, 100, 10, cp)

print(v.shape)
print(arr.nbytes, v.nbytes)
print(v.base)

The output of this example shows:

(1, 1, 91, 100)
8000 72800
<numpy.lib.stride_tricks.DummyArray object at 0x7f74ef128b38>
8000
(1, 1, 91, 100)
8000 72800
None

So it appears that when using as_strided, Cupy using the original array as a base. Does anyone have a suggestion on how to deal with this?

[newmanwang]

@kkotyk, Cupy's ndarray has __cuda_array_interface__ attribute that similar to numpy ndarray's __array_interface__. To avoid memory allocation as cupy's as_strided doese, as what numpy's as_strided did, make a DummyArray which has a __cuda_array_interface__, set desired strides and shape which make the sliding window view into this __cuda_array_interface__. As cp.asarray hornor strides and shape in the __cuda_array_interface__, so the new ndarray returned by cp.asarray(DummyArray(this_cuda_interface, base=original_array)) is almost the final real sliding window view that you can use in some cases(In other cases, for example, print this view lead to memory allocation ).