GPU Array Computing

Authors

Emily Costa

Created on

08/07/2019

The following are lessons learned during the exploration of implementing CuPy instead of numpy for GPU computing:

newaxis

Dimensions: CuPy does not have a newaxis function unlike NumPy. Instead of using new axis to add an additional dimension, you need to use cupy.expand_dims(). Also, note that CuPy does not lose a dimension during operations with vectors unlike numpy. So, adding another dimension is often unnecessary as there are no singular dimensions in CuPy. All vectors are converted into row vectors in numpy after being operated on, which can be dealt with by adding a new axis and converting back into a column vector for further matrix operations.

The following is an example of how numpy’s neawaxis function and how to use cupy’s expand_dims in its place:

numpy.newaxis

Import all necessary modules

In [1]: import numpy as np

1D array:

In [2]: arr = np.arange(5)

In [3]: arr.shape
Out[3]: (5,)

Make the 1D array becomes a row vector when an axis is inserted along 1st dimension

In [4]: row_vec = arr[np.newaxis, :]

In [5]: row_vec.shape
Out[5]: (1, 5)

Make the 1D array becomes a column vector when an axis is inserted along 1st dimension

In [6]: col_vec = arr[:, np.newaxis]

In [7]: col_vec.shape
Out[7]: (5, 1)

cupy.expand_dims

Import all necessary modules

In [1]: import cupy as cp

1D array

In [2]: cp_arr = cp.arange(5)

In [3]: cp_arr.shape
Out[3]: (5,)

Make the 1D array becomes a row vector when an axis is inserted along 1st dimension

In [4]: cp_row_vec = cp.expand_dims(cp_arr, axis=0)

In [5]: cp_row_vec.shape
Out[5]: (1, 5)

Make the 1D array becomes a column vector when an axis is inserted along 1st dimension

In [6]: cp_col_vec = cp.expand_dims(cp_arr, axis=1)

In [7]: cp_col_vec.shape
Out[7]: (5, 1)

append

CuPy does not have an append() function unlike NumPy. As a reminder - the append function in the NumPy appends values to the end of an array.

The following is an example of numpy’s append function and how to use cupy’s concatenate instead:

numpy.append

In [1]: x = np.array([1,2,3])

In [2]: y = [4,5,6]

In [3]: xy = np.append(x, y)

In [4]: xy
Out[4]: array([1,2,3,4,5,6])

cupy.concatenate

In [1]: cp_x = cp.array([1,2,3])

In [2]: cp_y = cp.array([4,5,6])

In [3]: cp_xy = cp.concatenate([cp_x,cp_y], axis=0)

In [4]: cp_xy
Out[4]: [1 2 3 4 5 6]