np.roll improvements

Two quick suggestions:

• add an inplace={False|True} (or perhaps just out, like for the ufuncs).
• support shifting along multiple axes simultaneously: either allow shift to be an array-like of length equal to the array’s ndim, or allow passing two sequences, shift and axis, of the same length (or allow both).

[link89]

Is there API that I can use an array as a shift register ? For example,

arr = np.arange(10)
new_data = np.zeros(4)
arr.shift(new_data)
# then arr should become: [4, 5, 6, 7, 8, 9, 0, 0, 0, 0] , the array is left shifted by 4 unit, and new_data is insert to the end of the array

This pattern is very common in real time signal process. I really hope numpy can have such API to reduce unnecessary memory copy.

[anntzer]

At least in the case where the shift is small compared to the size of the array (and I expect shifts by 1 to be quite common), roll could at least work by allocating a small array equal to the number of shifted elements as temporary storage and do the shifting in place.